Add files using upload-large-folder tool

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-023631c9-84c3-44a8-86f0-5eae46cd682c1758617643748-2025_09_23-10.54.11.633/source.csv

@@ -0,0 +1,10 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"genie.py",0,0,"from typing import Dict\n\nimport einops\nimport jax\nimport jax.numpy as jnp\nimport flax.nnx as nnx\nimport orbax.checkpoint as ocp\n\nfrom models.dynamics import DynamicsMaskGIT, DynamicsCausal\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\n\nclass Genie(nnx.Module):\n    """"""Genie model""""""\n\n    def __init__(\n        self,\n        in_dim: int,\n        tokenizer_dim: int,\n        tokenizer_ffn_dim: int,\n        latent_patch_dim: int,\n        num_patch_latents: int,\n        patch_size: int,\n        tokenizer_num_blocks: int,\n        tokenizer_num_heads: int,\n        lam_dim: int,\n        lam_ffn_dim: int,\n        latent_action_dim: int,\n        num_latent_actions: int,\n        lam_patch_size: int,\n        lam_num_blocks: int,\n        lam_num_heads: int,\n        lam_co_train: bool,\n        use_gt_actions: bool,\n        dyna_type: str,\n        dyna_dim: int,\n        dyna_ffn_dim: int,\n        dyna_num_blocks: int,\n        dyna_num_heads: int,\n        param_dtype: jnp.dtype,\n        dtype: jnp.dtype,\n        use_flash_attention: bool,\n        decode: bool,\n        rngs: nnx.Rngs,\n        dropout: float = 0.0,\n        mask_limit: float = 0.0,\n    ):\n        # --- Tokenizer ---\n        self.in_dim = in_dim\n        self.tokenizer_dim = tokenizer_dim\n        self.tokenizer_ffn_dim = tokenizer_ffn_dim\n        self.latent_patch_dim = latent_patch_dim\n        self.num_patch_latents = num_patch_latents\n        self.patch_size = patch_size\n        self.tokenizer_num_blocks = tokenizer_num_blocks\n        self.tokenizer_num_heads = tokenizer_num_heads\n        # --- LAM ---\n        self.lam_dim = lam_dim\n        self.lam_ffn_dim = lam_ffn_dim\n        self.latent_action_dim = latent_action_dim\n        self.num_latent_actions = num_latent_actions\n        self.lam_patch_size = lam_patch_size\n        self.lam_num_blocks = lam_num_blocks\n        self.lam_num_heads = lam_num_heads\n        self.lam_co_train = lam_co_train\n        self.use_gt_actions = use_gt_actions\n        # --- Dynamics ---\n        self.dyna_type = dyna_type\n        self.dyna_dim = dyna_dim\n        self.dyna_ffn_dim = dyna_ffn_dim\n        self.dyna_num_blocks = dyna_num_blocks\n        self.dyna_num_heads = dyna_num_heads\n        self.param_dtype = param_dtype\n        self.dtype = dtype\n        self.use_flash_attention = use_flash_attention\n        self.dropout = dropout\n        self.mask_limit = mask_limit\n        self.decode = decode\n\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            ffn_dim=self.tokenizer_ffn_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n            param_dtype=self.param_dtype,\n            dtype=self.dtype,\n            use_flash_attention=self.use_flash_attention,\n            rngs=rngs,\n        )\n        if self.use_gt_actions:\n            self.action_embed = nnx.Embed(\n                self.num_latent_actions, self.latent_action_dim, rngs=rngs\n            )\n            self.lam = None\n        else:\n            self.lam = LatentActionModel(\n                in_dim=self.in_dim,\n                model_dim=self.lam_dim,\n                ffn_dim=self.lam_ffn_dim,\n                latent_dim=self.latent_patch_dim,\n                num_latents=self.num_latent_actions,\n                patch_size=self.lam_patch_size,\n                num_blocks=self.lam_num_blocks,\n                num_heads=self.lam_num_heads,\n                dropout=0.0,\n                codebook_dropout=0.0,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                rngs=rngs,\n            )\n            self.action_embed = None\n        if self.dyna_type == ""maskgit"":\n            self.dynamics = DynamicsMaskGIT(\n                model_dim=self.dyna_dim,\n                ffn_dim=self.dyna_ffn_dim,\n                num_latents=self.num_patch_latents,\n                latent_action_dim=self.latent_action_dim,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                mask_limit=self.mask_limit,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                rngs=rngs,\n            )\n        elif self.dyna_type == ""causal"":\n            self.dynamics = DynamicsCausal(\n                model_dim=self.dyna_dim,\n                ffn_dim=self.dyna_ffn_dim,\n                num_latents=self.num_patch_latents,\n                latent_action_dim=self.latent_action_dim,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                decode=decode,\n                rngs=rngs,\n            )\n        else:\n            raise ValueError(f""Invalid dynamics type: {self.dyna_type}"")\n\n    def __call__(\n        self,\n        batch: Dict[str, jax.Array],\n        training: bool = True,\n    ) -> Dict[str, jax.Array]:\n        videos_BTHWC = batch[""videos""]\n        tokenizer_outputs = self.tokenizer.vq_encode(videos_BTHWC, training=False)\n        token_indices_BTN = tokenizer_outputs[""indices""]\n        latent_actions_BTm11L = None\n        action_embeddings_BTm11L = None\n        if self.use_gt_actions:\n            assert self.action_embed is not None\n            action_indices_E = None\n            action_embeddings_BT1L = self.action_embed(batch[""actions""]).reshape(\n                *batch[""actions""].shape[:2], 1, self.latent_action_dim\n            )\n            action_embeddings_BTm11L = action_embeddings_BT1L[:, 1:]\n        else:\n            assert self.lam is not None\n            lam_outputs = self.lam.vq_encode(videos_BTHWC, training=False)\n            z_q_BTm11L = lam_outputs[""z_q""]\n            action_indices_E = lam_outputs[""indices""]\n            latent_actions_BTm11L = jax.lax.cond(\n                self.lam_co_train,\n                lambda: z_q_BTm11L,\n                lambda: jax.lax.stop_gradient(z_q_BTm11L),\n            )\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(token_indices_BTN),\n            latent_actions=(\n                action_embeddings_BTm11L\n                if self.use_gt_actions\n                else latent_actions_BTm11L\n            ),\n        )\n        outputs[""mask_rng""] = batch[""rng""]\n        dyna_logits_BTNV, dyna_mask = self.dynamics(outputs, training)\n        outputs[""token_logits""] = dyna_logits_BTNV\n        outputs[""mask""] = dyna_mask\n        mle_indices_BTN = jnp.argmax(outputs[""token_logits""], axis=-1)\n        H, W = batch[""videos""].shape[2:4]\n        outputs[""recon""] = self.tokenizer.decode(mle_indices_BTN, (H, W))\n        if action_indices_E is not None:\n            outputs[""lam_indices""] = action_indices_E\n        return outputs\n\n    def sample(\n        self,\n        batch: Dict[str, jax.Array],\n        seq_len: int,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n        maskgit_steps: int = 25,\n    ) -> tuple[jax.Array, jax.Array]:\n        if self.dyna_type == ""maskgit"":\n            return self.sample_maskgit(\n                batch, seq_len, maskgit_steps, temperature, sample_argmax\n            )\n        elif self.dyna_type == ""causal"":\n            return self.sample_causal(batch, seq_len, temperature, sample_argmax)\n        else:\n            raise ValueError(f""Dynamics model type unknown: {self.dyna_type}"")\n\n    def sample_maskgit(\n        self,\n        batch: Dict[str, jax.Array],\n        seq_len: int,\n        steps: int = 25,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> tuple[jax.Array, jax.Array]:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by\n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size\n            T: number of input (conditioning) frames\n            N: number of patches per frame\n            M: model dimension\n            S: sequence length\n            H: height\n            W: width\n            E: B * (S - 1)\n            P: S * N\n        """"""\n        assert isinstance(self.dynamics, DynamicsMaskGIT)\n        # --- Encode videos and actions ---\n        videos_BTHWC = batch[""videos""]\n        tokenizer_out = self.tokenizer.vq_encode(videos_BTHWC, training=False)\n        token_idxs_BTN = tokenizer_out[""indices""]\n        B, T, N = token_idxs_BTN.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs_BTN.dtype)\n        token_idxs_BSN = jnp.concatenate([token_idxs_BTN, pad], axis=1)\n        init_logits_BSNV = jnp.zeros(\n            shape=(*token_idxs_BSN.shape, self.num_patch_latents)\n        )\n        if self.use_gt_actions:\n            assert self.action_embed is not None\n            latent_actions_BT1L = self.action_embed(batch[""actions""]).reshape(\n                *batch[""actions""].shape[:2], 1, self.latent_action_dim\n            )\n            latent_actions_BTm11L = latent_actions_BT1L[:, 1:]\n            action_tokens_EL = latent_actions_BTm11L.reshape(-1, self.latent_action_dim)\n        else:\n            assert self.lam is not None\n            latent_actions_E = batch[""latent_actions""]\n            action_tokens_EL = self.lam.vq.get_codes(latent_actions_E)\n\n        # --- Extract submodule state ---\n        dynamics_state = nnx.state(self.dynamics)\n\n        @nnx.scan(in_axes=(nnx.Carry, 0), out_axes=nnx.Carry)\n        def maskgit_step_fn(\n            carry: tuple[jax.Array, jax.Array, jax.Array, jax.Array, jax.Array],\n            step: jax.Array,\n        ) -> tuple[jax.Array, jax.Array, jax.Array, jax.Array, jax.Array]:\n            rng, token_idxs_BSN, logits_BSNV, mask_BSN, action_tokens_EL = carry\n            S, N = token_idxs_BSN.shape[1:]\n            L = action_tokens_EL.shape[-1]\n\n            # We need to reconstruct the submodule inside scan body to prevent trace context mismatches\n            dynamics_maskgit = DynamicsMaskGIT(\n                model_dim=self.dyna_dim,\n                ffn_dim=self.dyna_ffn_dim,\n                num_latents=self.num_patch_latents,\n                latent_action_dim=self.latent_action_dim,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                mask_limit=self.mask_limit,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                rngs=nnx.Rngs(0),\n            )\n            nnx.update(dynamics_maskgit, dynamics_state)\n\n            # --- Construct + encode video ---\n            vid_embed_BSNM = dynamics_maskgit.patch_embed(token_idxs_BSN)\n            mask_token_111M = dynamics_maskgit.mask_token.value\n            mask_expanded_BSN1 = mask_BSN[..., None]\n            vid_embed_BSNM = jnp.where(\n                mask_expanded_BSN1, mask_token_111M, vid_embed_BSNM\n            )\n\n            # --- Predict transition ---\n            action_tokens_BSm1L = jnp.reshape(action_tokens_EL, (B, S - 1, L))\n            act_embed_BSm1M = dynamics_maskgit.action_up(action_tokens_BSm1L)\n            act_embed_BSM = jnp.pad(act_embed_BSm1M, ((0, 0), (1, 0), (0, 0)))\n            act_embed_BS1M = jnp.reshape(\n                act_embed_BSM, (B, S, 1, act_embed_BSM.shape[-1])\n            )\n            vid_embed_BSNM += act_embed_BS1M\n            unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (steps * 2))\n            step_temp = temperature * (1.0 - unmasked_ratio)\n            final_logits_BSNV = dynamics_maskgit.transformer(vid_embed_BSNM) / step_temp\n\n            # --- Sample new tokens for final frame ---\n            if sample_argmax:\n                sampled_token_idxs_BSN = jnp.argmax(final_logits_BSNV, axis=-1)\n            else:\n                rng, _rng = jax.random.split(rng)\n                sampled_token_idxs_BSN = jax.random.categorical(_rng, final_logits_BSNV)\n            gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n            final_token_probs_BSN = gather_fn(\n                jax.nn.softmax(final_logits_BSNV), sampled_token_idxs_BSN\n            )\n            final_token_probs_BSN += ~mask_BSN\n            # Update masked tokens and logits only\n            token_idxs_BSN = jnp.where(mask_BSN, sampled_token_idxs_BSN, token_idxs_BSN)\n            logits_BSNV = jnp.where(\n                jnp.expand_dims(mask_BSN, -1), final_logits_BSNV, logits_BSNV\n            )\n\n            # --- Update mask ---\n            num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n            final_token_probs_flat_BP = einops.rearrange(\n                final_token_probs_BSN, ""b s n -> b (s n)""\n            )\n            idx_mask_P = (\n                jnp.arange(final_token_probs_flat_BP.shape[-1])\n                <= N - num_unmasked_tokens\n            )\n            sorted_idxs_BP = jnp.argsort(final_token_probs_flat_BP, axis=-1)\n            mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask_P))\n            mask_flat_BP = einops.rearrange(mask_BSN, ""b s n -> b (s n)"")\n            new_mask_flat_BP = mask_update_fn(mask_flat_BP, sorted_idxs_BP)\n            new_mask_BSN = einops.rearrange(new_mask_flat_BP, ""b (s n) -> b s n"", n=N)\n\n            new_carry = (\n                rng,\n                token_idxs_BSN,\n                logits_BSNV,\n                new_mask_BSN,\n                action_tokens_EL,\n            )\n            return new_carry\n\n        @nnx.scan(in_axes=(nnx.Carry, 0), out_axes=nnx.Carry)\n        def generation_step_fn(\n            carry: tuple[jax.Array, jax.Array, jax.Array], step_t: jax.Array\n        ) -> tuple[jax.Array, jax.Array, jax.Array]:\n            rng, current_token_idxs_BSN, current_logits_BSNV = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # Mask current frame (i.e., t == step_t)\n            mask_S = jnp.arange(seq_len) == step_t\n            mask_BSN = jnp.broadcast_to(mask_S[None, :, None], (B, seq_len, N)).astype(\n                bool\n            )\n            masked_token_idxs_BSN = current_token_idxs_BSN * ~mask_BSN\n            masked_logits_BSNV = current_logits_BSNV * jnp.expand_dims(~mask_BSN, -1)\n\n            # --- Initialize and run MaskGIT loop ---\n            init_carry_maskgit = (\n                step_rng,\n                masked_token_idxs_BSN,\n                masked_logits_BSNV,\n                mask_BSN,\n                action_tokens_EL,\n            )\n            final_carry_maskgit = maskgit_step_fn(init_carry_maskgit, jnp.arange(steps))\n            updated_token_idxs_BSN = final_carry_maskgit[1]\n            updated_logits_BSNV = final_carry_maskgit[2]\n            new_carry = (rng, updated_token_idxs_BSN, updated_logits_BSNV)\n            return new_carry\n\n        # --- Run the autoregressive generation using jax.lax.scan ---\n        initial_carry = (batch[""rng""], token_idxs_BSN, init_logits_BSNV)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry = generation_step_fn(initial_carry, timesteps_to_scan)\n        final_token_idxs_BSN = final_carry[1]\n        final_logits_BSNV = final_carry[2]\n\n        # --- Decode all tokens at once at the end ---\n        H, W = batch[""videos""].shape[2:4]\n        final_frames_BSHWC = self.tokenizer.decode(\n            final_token_idxs_BSN,\n            video_hw=(H, W),\n        )\n        return final_frames_BSHWC, final_logits_BSNV\n\n    def sample_causal(\n        self,\n        batch: Dict[str, jax.Array],\n        seq_len: int,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> tuple[jax.Array, jax.Array]:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by\n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size\n            T: number of input (conditioning) frames\n            N: number of patches per frame\n            M: model dimension\n            S: sequence length\n            H: height\n            W: width\n            E: B * (S - 1)\n        """"""\n        assert isinstance(self.dynamics, DynamicsCausal)\n        # --- Encode videos and actions ---\n        videos_BTHWC = batch[""videos""]\n        tokenizer_out = self.tokenizer.vq_encode(videos_BTHWC, training=False)\n        token_idxs_BTN = tokenizer_out[""indices""]\n        B, T, N = token_idxs_BTN.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs_BTN.dtype)\n        token_idxs_BSN = jnp.concatenate([token_idxs_BTN, pad], axis=1)\n        logits_BSNV = jnp.zeros((*token_idxs_BSN.shape, self.num_patch_latents))\n        dynamics_state = nnx.state(self.dynamics)\n\n        if self.use_gt_actions:\n            assert self.action_embed is not None\n            latent_actions_BT1L = self.action_embed(batch[""actions""]).reshape(\n                *batch[""actions""].shape[:2], 1, self.latent_action_dim\n            )\n            latent_actions_BTm11L = latent_actions_BT1L[:, 1:]\n            action_tokens_EL = latent_actions_BTm11L.reshape(-1, self.latent_action_dim)\n        else:\n            assert self.lam is not None\n            latent_actions_E = batch[""latent_actions""]\n            action_tokens_EL = self.lam.vq.get_codes(latent_actions_E)\n\n        @nnx.scan(in_axes=(nnx.Carry, 0), out_axes=nnx.Carry)\n        def causal_step_fn(\n            carry: tuple[jax.Array, jax.Array, jax.Array, jax.Array, jax.Array],\n            step_n: jax.Array,\n        ) -> tuple[jax.Array, jax.Array, jax.Array, jax.Array, jax.Array]:\n            rng, token_idxs_BSN, logits_BSNV, action_tokens_EL, step_t = carry\n            S, N = token_idxs_BSN.shape[1:]\n            L = action_tokens_EL.shape[-1]\n\n            # We need to reconstruct the submodule inside scan body to prevent trace context mismatches\n            dynamics_causal = DynamicsCausal(\n                model_dim=self.dyna_dim,\n                ffn_dim=self.dyna_ffn_dim,\n                num_latents=self.num_patch_latents,\n                latent_action_dim=self.latent_action_dim,\n                num_blocks=self.dyna_num_blocks,\n                num_heads=self.dyna_num_heads,\n                dropout=self.dropout,\n                param_dtype=self.param_dtype,\n                dtype=self.dtype,\n                use_flash_attention=self.use_flash_attention,\n                decode=self.decode,\n                rngs=nnx.Rngs(0),\n            )\n            nnx.update(dynamics_causal, dynamics_state)\n\n            # --- Construct + encode video ---\n            vid_embed_BSNM = dynamics_causal.patch_embed(token_idxs_BSN)\n\n            # --- Predict transition ---\n            action_tokens_BSm1L = jnp.reshape(action_tokens_EL, (B, S - 1, L))\n            act_embed_BSm1M = dynamics_causal.action_up(action_tokens_BSm1L)\n            act_embed_BSM = jnp.pad(act_embed_BSm1M, ((0, 0), (1, 0), (0, 0)))\n            act_embed_BS1M = jnp.reshape(\n                act_embed_BSM, (B, S, 1, act_embed_BSM.shape[-1])\n            )\n            vid_embed_BSNp1M = jnp.concatenate([act_embed_BS1M, vid_embed_BSNM], axis=2)\n            final_logits_BTNp1V = (\n                dynamics_causal.transformer(vid_embed_BSNp1M, (step_t, step_n))\n                / temperature\n            )\n            final_logits_BV = final_logits_BTNp1V[:, step_t, step_n, :]\n\n            # --- Sample new tokens for final frame ---\n            if sample_argmax:\n                sampled_token_idxs_B = jnp.argmax(final_logits_BV, axis=-1)\n            else:\n                rng, _rng = jax.random.split(rng)\n                sampled_token_idxs_B = jax.random.categorical(_rng, final_logits_BV)\n            # Update next tokens only\n            token_idxs_BSN = token_idxs_BSN.at[:, step_t, step_n].set(\n                sampled_token_idxs_B\n            )\n            logits_BSNV = logits_BSNV.at[:, step_t, step_n].set(final_logits_BV)\n\n            new_carry = (rng, token_idxs_BSN, logits_BSNV, action_tokens_EL, step_t)\n            return new_carry\n\n        @nnx.scan(in_axes=(nnx.Carry, 0), out_axes=nnx.Carry)\n        def generation_step_fn(\n            carry: tuple[jax.Array, jax.Array, jax.Array], step_t: jax.Array\n        ) -> tuple[jax.Array, jax.Array, jax.Array]:\n            rng, current_token_idxs_BSN, current_logits_BSNV = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # --- Initialize and run causal loop ---\n            init_carry_causal = (\n                step_rng,\n                current_token_idxs_BSN,\n                current_logits_BSNV,\n                action_tokens_EL,\n                step_t,\n            )\n            final_carry_causal = causal_step_fn(init_carry_causal, jnp.arange(N))\n            updated_token_idxs_BSN = final_carry_causal[1]\n            updated_logits_BSNV = final_carry_causal[2]\n            new_carry = (rng, updated_token_idxs_BSN, updated_logits_BSNV)\n            return new_carry\n\n        # --- Run the autoregressive generation using jax.lax.scan ---\n        initial_carry = (batch[""rng""], token_idxs_BSN, logits_BSNV)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry = generation_step_fn(initial_carry, timesteps_to_scan)\n        final_token_idxs_BSN = final_carry[1]\n        final_logits_BSNV = final_carry[2]\n\n        # --- Decode all tokens at once at the end ---\n        H, W = batch[""videos""].shape[2:4]\n        final_frames_BSHWC = self.tokenizer.decode(\n            final_token_idxs_BSN,\n            video_hw=(H, W),\n        )\n        return final_frames_BSHWC, final_logits_BSNV\n\n    def vq_encode(self, batch: Dict[str, jax.Array], training: bool) -> jax.Array:\n        # --- Preprocess videos ---\n        assert self.lam is not None\n        video_BTHWC = batch[""videos""]\n        lam_output: Dict[str, jax.Array] = self.lam.vq_encode(\n            video_BTHWC, training=training\n        )\n        lam_indices_E = lam_output[""indices""]\n        return lam_indices_E\n\n\n# FIXME (f.srambical): add conversion script for old checkpoints\ndef restore_genie_components(\n    optimizer: nnx.Optimizer,\n    sharding: jax.sharding.NamedSharding,\n    rng: jax.Array,\n    args,\n) -> nnx.Optimizer:\n    """"""Restore pre-trained Genie components""""""\n    rng_tokenizer, rng_lam = jax.random.split(rng)\n    rngs_tokenizer = nnx.Rngs(rng_tokenizer)\n    rngs_lam = nnx.Rngs(rng_lam)\n\n    tx = optimizer.tx\n    model = optimizer.model\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        step_format_fixed_length=6,\n    )\n    tokenizer_checkpoint_manager = ocp.CheckpointManager(\n        directory=args.tokenizer_checkpoint,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    dummy_tokenizer = TokenizerVQVAE(\n        in_dim=args.image_channels,\n        model_dim=args.tokenizer_dim,\n        ffn_dim=args.tokenizer_ffn_dim,\n        latent_dim=args.latent_patch_dim,\n        num_latents=args.num_patch_latents,\n        patch_size=args.patch_size,\n        num_blocks=args.tokenizer_num_blocks,\n        num_heads=args.tokenizer_num_heads,\n        dropout=args.dropout,\n        codebook_dropout=args.dropout,\n        param_dtype=args.param_dtype,\n        dtype=args.dtype,\n        use_flash_attention=args.use_flash_attention,\n        rngs=rngs_tokenizer,\n    )\n    dummy_tokenizer_optimizer = nnx.Optimizer(dummy_tokenizer, tx)\n    dummy_tokenizer_optimizer_state = nnx.state(dummy_tokenizer_optimizer)\n    abstract_sharded_tokenizer_optimizer_state = _create_abstract_sharded_pytree(\n        dummy_tokenizer_optimizer_state, sharding\n    )\n    restored_tokenizer = tokenizer_checkpoint_manager.restore(\n        step=tokenizer_checkpoint_manager.latest_step(),\n        args=ocp.args.Composite(\n            model_state=ocp.args.PyTreeRestore(  # type: ignore\n                abstract_sharded_tokenizer_optimizer_state  # type: ignore\n            ),\n        ),\n    )[""model_state""]\n    nnx.update(dummy_tokenizer_optimizer.model, restored_tokenizer.model)\n    model.tokenizer = dummy_tokenizer_optimizer.model\n    tokenizer_checkpoint_manager.close()\n\n    if args.lam_checkpoint:\n        lam_checkpoint_manager = ocp.CheckpointManager(\n            directory=args.lam_checkpoint,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        dummy_lam = LatentActionModel(\n            in_dim=args.image_channels,\n            model_dim=args.lam_dim,\n            ffn_dim=args.lam_ffn_dim,\n            latent_dim=args.latent_patch_dim,\n            num_latents=args.num_latent_actions,\n            patch_size=args.lam_patch_size,\n            num_blocks=args.lam_num_blocks,\n            num_heads=args.lam_num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs_lam,\n        )\n        dummy_lam_optimizer = nnx.Optimizer(dummy_lam, tx)\n        dummy_lam_optimizer_state = nnx.state(dummy_lam_optimizer)\n        abstract_sharded_lam_optimizer_state = _create_abstract_sharded_pytree(\n            dummy_lam_optimizer_state, sharding\n        )\n        restored_lam_optimizer = lam_checkpoint_manager.restore(\n            step=lam_checkpoint_manager.latest_step(),\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(  # type: ignore\n                    abstract_sharded_lam_optimizer_state  # type: ignore\n                ),\n            ),\n        )[""model_state""]\n        nnx.update(dummy_lam_optimizer.model, restored_lam_optimizer.model)\n        model.lam = dummy_lam_optimizer.model\n        # Remove the LAM decoder to save memory and avoid unnecessary computation.\n        del model.lam.decoder\n        lam_checkpoint_manager.close()\n\n    # Reinitialize the optimizer states\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer\n\n\ndef _create_abstract_sharded_pytree(\n    pytree_template: nnx.GraphState, sharding_spec: jax.sharding.NamedSharding\n) -> jax.Array:\n    """"""Replaces arrays in a pytree with ShapeDtypeStructs having the given sharding.""""""\n\n    def map_fn(leaf_template):\n        if hasattr(leaf_template, ""shape"") and hasattr(leaf_template, ""dtype""):\n            return jax.ShapeDtypeStruct(\n                leaf_template.shape, leaf_template.dtype, sharding=sharding_spec\n            )\n        return leaf_template\n\n    return jax.tree_util.tree_map(map_fn, pytree_template)\n",python,tab
+2,136,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:54:11 AM [info] Activating crowd-code\n10:54:11 AM [info] Recording started\n10:54:11 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,172,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"10:54:11 AM [info] Git repository found\n10:54:11 AM [info] Git provider initialized successfully\n10:54:11 AM [info] Initial git state: [object Object]\n",Log,content
+4,4275,"TERMINAL",0,0,"",,terminal_command
+5,8212,"genie.py",0,0,"",python,tab
+6,9295,"TERMINAL",0,0,"ls /fast/project/HFMI_SynergyUnit/jafar_ws/data/breakout/train/",,terminal_command
+7,9350,"TERMINAL",0,0,"]633;Cdata_0000.array_record  data_0084.array_record  data_0168.array_record  data_0252.array_record  data_0336.array_record  data_0420.array_record  data_0504.array_record  data_0588.array_record\r\ndata_0001.array_record  data_0085.array_record  data_0169.array_record  data_0253.array_record  data_0337.array_record  data_0421.array_record  data_0505.array_record  data_0589.array_record\r\ndata_0002.array_record  data_0086.array_record  data_0170.array_record  data_0254.array_record  data_0338.array_record  data_0422.array_record  data_0506.array_record  data_0590.array_record\r\ndata_0003.array_record  data_0087.array_record  data_0171.array_record  data_0255.array_record  data_0339.array_record  data_0423.array_record  data_0507.array_record  data_0591.array_record\r\ndata_0004.array_record  data_0088.array_record  data_0172.array_record  data_0256.array_record  data_0340.array_record  data_0424.array_record  data_0508.array_record  data_0592.array_record\r\ndata_0005.array_record  data_0089.array_record  data_0173.array_record  data_0257.array_record  data_0341.array_record  data_0425.array_record  data_0509.array_record  data_0593.array_record\r\ndata_0006.array_record  data_0090.array_record  data_0174.array_record  data_0258.array_record  data_0342.array_record  data_0426.array_record  data_0510.array_record  data_0594.array_record\r\ndata_0007.array_record  data_0091.array_record  data_0175.array_record  data_0259.array_record  data_0343.array_record  data_0427.array_record  data_0511.array_record  data_0595.array_record\r\ndata_0008.array_record  data_0092.array_record  data_0176.array_record  data_0260.array_record  data_0344.array_record  data_0428.array_record  data_0512.array_record  data_0596.array_record\r\ndata_0009.array_record  data_0093.array_record  data_0177.array_record  data_0261.array_record  data_0345.array_record  data_0429.array_record  data_0513.array_record  data_0597.array_record\r\ndata_0010.array_record  data_0094.array_record  data_0178.array_record  data_0262.array_record  data_0346.array_record  data_0430.array_record  data_0514.array_record  data_0598.array_record\r\ndata_0011.array_record  data_0095.array_record  data_0179.array_record  data_0263.array_record  data_0347.array_record  data_0431.array_record  data_0515.array_record  data_0599.array_record\r\ndata_0012.array_record  data_0096.array_record  data_0180.array_record  data_0264.array_record  data_0348.array_record  data_0432.array_record  data_0516.array_record  data_0600.array_record\r\ndata_0013.array_record  data_0097.array_record  data_0181.array_record  data_0265.array_record  data_0349.array_record  data_0433.array_record  data_0517.array_record  data_0601.array_record\r\ndata_0014.array_record  data_0098.array_record  data_0182.array_record  data_0266.array_record  data_0350.array_record  data_0434.array_record  data_0518.array_record  data_0602.array_record\r\ndata_0015.array_record  data_0099.array_record  data_0183.array_record  data_0267.array_record  data_0351.array_record  data_0435.array_record  data_0519.array_record  data_0603.array_record\r\ndata_0016.array_record  data_0100.array_record  data_0184.array_record  data_0268.array_record  data_0352.array_record  data_0436.array_record  data_0520.array_record  data_0604.array_record\r\ndata_0017.array_record  data_0101.array_record  data_0185.array_record  data_0269.array_record  data_0353.array_record  data_0437.array_record  data_0521.array_record  data_0605.array_record\r\ndata_0018.array_record  data_0102.array_record  data_0186.array_record  data_0270.array_record  data_0354.array_record  data_0438.array_record  data_0522.array_record  data_0606.array_record\r\ndata_0019.array_record  data_0103.array_record  data_0187.array_record  data_0271.array_record  data_0355.array_record  data_0439.array_record  data_0523.array_record  data_0607.array_record\r\ndata_0020.array_record  data_0104.array_record  data_0188.array_record  data_0272.array_record  data_0356.array_record  data_0440.array_record  data_0524.array_record  data_0608.array_record\r\ndata_0021.array_record  data_0105.array_record  data_0189.array_record  data_0273.array_record  data_0357.array_record  data_0441.array_record  data_0525.array_record  data_0609.array_record\r\ndata_0022.array_record  data_0106.array_record  data_0190.array_record  data_0274.array_record  data_0358.array_record  data_0442.array_record  data_0526.array_record  data_0610.array_record\r\ndata_0023.array_record  data_0107.array_record  data_0191.array_record  data_0275.array_record  data_0359.array_record  data_0443.array_record  data_0527.array_record  data_0611.array_record\r\ndata_0024.array_record  data_0108.array_record  data_0192.array_record  data_0276.array_record  data_0360.array_record  data_0444.array_record  data_0528.array_record  data_0612.array_record\r\ndata_0025.array_record  data_0109.array_record  data_0193.array_record  data_0277.array_record  data_0361.array_record  data_0445.array_record  data_0529.array_record  data_0613.array_record\r\ndata_0026.array_record  data_0110.array_record  data_0194.array_record  data_0278.array_record  data_0362.array_record  data_0446.array_record  data_0530.array_record  data_0614.array_record\r\ndata_0027.array_record  data_0111.array_record  data_0195.array_record  data_0279.array_record  data_0363.array_record  data_0447.array_record  data_0531.array_record  data_0615.array_record\r\ndata_0028.array_record  data_0112.array_record  data_0196.array_record  data_0280.array_record  data_0364.array_record  data_0448.array_record  data_0532.array_record  data_0616.array_record\r\ndata_0029.array_record  data_0113.array_record  data_0197.array_record  data_0281.array_record  data_0365.array_record  data_0449.array_record  data_0533.array_record  data_0617.array_record\r\ndata_0030.array_record  data_0114.array_record  data_0198.array_record  data_0282.array_record  data_0366.array_record  data_0450.array_record  data_0534.array_record  data_0618.array_record\r\ndata_0031.array_record  data_0115.array_record  data_0199.array_record  data_0283.array_record  data_0367.array_record  data_0451.array_record  data_0535.array_record  data_0619.array_record\r\ndata_0032.array_record  data_0116.array_record  data_0200.array_record  data_0284.array_record  data_0368.array_record  data_0452.array_record  data_0536.array_record  data_0620.array_record\r\ndata_0033.array_record  data_0117.array_record  data_0201.array_record  data_0285.array_record  data_0369.array_record  data_0453.array_record  data_0537.array_record  data_0621.array_record\r\ndata_0034.array_record  data_0118.array_record  data_0202.array_record  data_0286.array_record  data_0370.array_record  data_0454.array_record  data_0538.array_record  data_0622.array_record\r\ndata_0035.array_record  data_0119.array_record  data_0203.array_record  data_0287.array_record  data_0371.array_record  data_0455.array_record  data_0539.array_record  data_0623.array_record\r\ndata_0036.array_record  data_0120.array_record  data_0204.array_record  data_0288.array_record  data_0372.array_record  data_0456.array_record  data_0540.array_record  data_0624.array_record\r\ndata_0037.array_record  data_0121.array_record  data_0205.array_record  data_0289.array_record  data_0373.array_record  data_0457.array_record  data_0541.array_record  data_0625.array_record\r\ndata_0038.array_record  data_0122.array_record  data_0206.array_record  data_0290.array_record  data_0374.array_record  data_0458.array_record  data_0542.array_record  data_0626.array_record\r\ndata_0039.array_record  data_0123.array_record  data_0207.array_record  data_0291.array_record  data_0375.array_record  data_0459.array_record  data_0543.array_record  data_0627.array_record\r\ndata_0040.array_record  data_0124.array_record  data_0208.array_record  data_0292.array_record  data_0376.array_record  data_0460.array_record  data_0544.array_record  data_0628.array_record\r\ndata_0041.array_record  data_0125.array_record  data_0209.array_record  data_0293.array_record  data_0377.array_record  data_0461.array_record  data_0545.array_record  data_0629.array_record\r\ndata_0042.array_record  data_0126.array_record  data_0210.array_record  data_0294.array_record  data_0378.array_record  data_0462.array_record  data_0546.array_record  data_0630.array_record\r\ndata_0043.array_record  data_0127.array_record  data_0211.array_record  data_0295.array_record  data_0379.array_record  data_0463.array_record  data_0547.array_record  data_0631.array_record\r\ndata_0044.array_record  data_0128.array_record  data_0212.array_record  data_0296.array_record  data_0380.array_record  data_0464.array_record  data_0548.array_record  data_0632.array_record\r\ndata_0045.array_record  data_0129.array_record  data_0213.array_record  data_0297.array_record  data_0381.array_record  data_0465.array_record  data_0549.array_record  data_0633.array_record\r\ndata_0046.array_record  data_0130.array_record  data_0214.array_record  data_0298.array_record  data_0382.array_record  data_0466.array_record  data_0550.array_record  data_0634.array_record\r\ndata_0047.array_record  data_0131.array_record  data_0215.array_record  data_0299.array_record  data_0383.array_record  data_0467.array_record  data_0551.array_record  data_0635.array_record\r\ndata_0048.array_record  data_0132.array_record  data_0216.array_record  data_0300.array_record  data_0384.array_record  data_0468.array_record  data_0552.array_record  data_0636.array_record\r\ndata_0049.array_record  data_0133.array_record  data_0217.array_record  data_0301.array_record  data_0385.array_record  data_0469.array_record  data_0553.array_record  data_0637.array_record\r\ndata_0050.array_record  data_0134.array_record  data_0218.array_record  data_0302.array_record  data_0386.array_record  data_0470.array_record  data_0554.array_record  data_0638.array_record\r\ndata_0051.array_record  data_0135.array_record  data_0219.array_record  data_0303.array_record  data_0387.array_record  data_0471.array_record  data_0555.array_record  data_0639.array_record\r\ndata_0052.array_record  data_0136.array_record  data_0220.array_record  data_0304.array_record  data_0388.array_record  data_0472.array_record  data_0556.array_record  data_0640.array_record\r\ndata_0053.array_record  data_0137.array_record  data_0221.array_record  data_0305.array_record  data_0389.array_record  data_0473.array_record  data_0557.array_record  data_0641.array_record\r\ndata_0054.array_record  data_0138.array_record  data_0222.array_record  data_0306.array_record  data_0390.array_record  data_0474.array_record  data_0558.array_record  data_0642.array_record\r\ndata_0055.array_record  data_0139.array_record  data_0223.array_record  data_0307.array_record  data_0391.array_record  data_0475.array_record  data_0559.array_record  data_0643.array_record\r\ndata_0056.array_record  data_0140.array_record  data_0224.array_record  data_0308.array_record  data_0392.array_record  data_0476.array_record  data_0560.array_record  data_0644.array_record\r\ndata_0057.array_record  data_0141.array_record  data_0225.array_record  data_0309.array_record  data_0393.array_record  data_0477.array_record  data_0561.array_record  data_0645.array_record\r\ndata_0058.array_record  data_0142.array_record  data_0226.array_record  data_0310.array_record  data_0394.array_record  data_0478.array_record  data_0562.array_record  data_0646.array_record\r\ndata_0059.array_record  data_0143.array_record  data_0227.array_record  data_0311.array_record  data_0395.array_record  data_0479.array_record  data_0563.array_record  data_0647.array_record\r\ndata_0060.array_record  data_0144.array_record  data_0228.array_record  data_0312.array_record  data_0396.array_record  data_0480.array_record  data_0564.array_record  data_0648.array_record\r\ndata_0061.array_record  data_0145.array_record  data_0229.array_record  data_0313.array_record  data_0397.array_record  data_0481.array_record  data_0565.array_record  data_0649.array_record\r\ndata_0062.array_record  data_0146.array_record  data_0230.array_record  data_0314.array_record  data_0398.array_record  data_0482.array_record  data_0566.array_record  data_0650.array_record\r\ndata_0063.array_record  data_0147.array_record  data_0231.array_record  data_0315.array_record  data_0399.array_record  data_0483.array_record  data_0567.array_record  data_0651.array_record\r\ndata_0064.array_record  data_0148.array_record  data_0232.array_record  data_0316.array_record  data_0400.array_record  data_0484.array_record  data_0568.array_record  data_0652.array_record\r\ndata_0065.array_record  data_0149.array_record  data_0233.array_record  data_0317.array_record  data_0401.array_record  data_0485.array_record  data_0569.array_record  data_0653.array_record\r\ndata_0066.array_record  data_0150.array_record  data_0234.array_record  data_0318.array_record  data_0402.array_record  data_0486.array_record  data_0570.array_record  data_0654.array_record\r\ndata_0067.array_record  data_0151.array_record  data_0235.array_record  data_0319.array_record  data_0403.array_record  data_0487.array_record  data_0571.array_record  data_0655.array_record\r\ndata_0068.array_record  data_0152.array_record  data_0236.array_record  data_0320.array_record  data_0404.array_record  data_0488.array_record  data_0572.array_record  data_0656.array_record\r\ndata_0069.array_record  data_0153.array_record  data_0237.array_record  data_0321.array_record  data_0405.array_record  data_0489.array_record  data_0573.array_record  data_0657.array_record\r\ndata_0070.array_record  data_0154.array_record  data_0238.array_record  data_0322.array_record  data_0406.array_record  data_0490.array_record  data_0574.array_record  data_0658.array_record\r\ndata_0071.array_record  data_0155.array_record  data_0239.array_record  data_0323.array_record  data_0407.array_record  data_0491.array_record  data_0575.array_record  data_0659.array_record\r\ndata_0072.array_record  data_0156.array_record  data_0240.array_record  data_0324.array_record  data_0408.array_record  data_0492.array_record  data_0576.array_record  data_0660.array_record\r\ndata_0073.array_record  data_0157.array_record  data_0241.array_record  data_0325.array_record  data_0409.array_record  data_0493.array_record  data_0577.array_record  data_0661.array_record\r\ndata_0074.array_record  data_0158.array_record  data_0242.array_record  data_0326.array_record  data_0410.array_record  data_0494.array_record  data_0578.array_record  data_0662.array_record\r\ndata_0075.array_record  data_0159.array_record  data_0243.array_record  data_0327.array_record  data_0411.array_record  data_0495.array_record  data_0579.array_record  data_0663.array_record\r\ndata_0076.array_record  data_0160.array_record  data_0244.array_record  data_0328.array_record  data_0412.array_record  data_0496.array_record  data_0580.array_record  data_0664.array_record\r\ndata_0077.array_record  data_0161.array_record  data_0245.array_record  data_0329.array_record  data_0413.array_record  data_0497.array_record  data_0581.array_record\r\ndata_0078.array_record  data_0162.array_record  data_0246.array_record  data_0330.array_record  data_0414.array_record  data_0498.array_record  data_0582.array_record\r\ndata_0079.array_record  data_0163.array_record  data_0247.array_record  data_0331.array_record  data_0415.array_record  data_0499.array_record  data_0583.array_record\r\ndata_0080.array_record  data_0164.array_record  data_0248.array_record  data_0332.array_record  data_0416.array_record  data_0500.array_record  data_0584.array_record\r\ndata_0081.array_record  data_0165.array_record  data_0249.array_record  data_0333.array_record  data_0417.array_record  data_0501.array_record  data_0585.array_record\r\ndata_0082.array_record  data_0166.array_record  data_0250.array_record  data_0334.array_record  data_0418.array_record  data_0502.array_record  data_0586.array_record\r\ndata_0083.array_record  data_0167.array_record  data_0251.array_record  data_0335.array_record  data_0419.array_record  data_0503.array_record  data_0587.array_record\r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+8,10583,"TERMINAL",0,0,"",,terminal_command
+9,118684,"TERMINAL",0,0,"",,terminal_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-0692028e-98ba-48e2-9ebc-ba3f1c4cd5a21759257473788-2025_09_30-20.38.00.177/source.csv

@@ -0,0 +1,19 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,169,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:38:00 PM [info] Activating crowd-code\n8:38:00 PM [info] Recording started\n8:38:00 PM [info] Initializing git provider using file system watchers...\n8:38:00 PM [info] Git repository found\n8:38:00 PM [info] Git provider initialized successfully\n8:38:00 PM [info] Initial git state: [object Object]\n",Log,tab
+3,5360,"TERMINAL",0,0,"",,terminal_command
+4,12136,"TERMINAL",0,0,"",,terminal_command
+5,69015,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+6,78165,"TERMINAL",0,0,"",,terminal_focus
+7,86274,"TERMINAL",0,0,"git pull",,terminal_command
+8,86329,"TERMINAL",0,0,"]633;C",,terminal_output
+9,87717,"TERMINAL",0,0,"hint: You have divergent branches and need to specify how to reconcile them.\r\nhint: You can do so by running one of the following commands sometime before\r\nhint: your next pull:\r\nhint:\r\nhint:   git config pull.rebase false  # merge\r\nhint:   git config pull.rebase true   # rebase\r\nhint:   git config pull.ff only       # fast-forward only\r\nhint:\r\nhint: You can replace ""git config"" with ""git config --global"" to set a default\r\nhint: preference for all repositories. You can also pass --rebase, --no-rebase,\r\nhint: or --ff-only on the command line to override the configured default per\r\nhint: invocation.\r\nfatal: Need to specify how to reconcile divergent branches.\r\n]0;franz.srambical@hai-login1:~/jafar/slurm",,terminal_output
+10,95773,"TERMINAL",0,0,"git config pull.rebase false",,terminal_command
+11,95780,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login1:~/jafar/slurm",,terminal_output
+12,96798,"TERMINAL",0,0,"git pull",,terminal_command
+13,96850,"TERMINAL",0,0,"]633;C",,terminal_output
+14,105808,"TERMINAL",0,0,"hint: Waiting for your editor to close the file... [?1049h[>4;2m[?1h=[?2004h[?1004h[?12h[?12l[?25l""~/jafar/slurm/.git/MERGE_MSG"" 6L, 273B▽  Pzz\[0%m           [>c]10;?]11;?Merge branch 'main' of github.com:p-doom/slurm\r\n# Please enter a commit message to explain why this merge is necessary,# especially if it merges an updated upstream into a topic branch.#\r\n# Lines starting with '#' will be ignored, and an empty message aborts\r\n# the commit.\r\n~                                                                                                                                     ~                                                                                                                                     ~                                                                                                                                     ~                                                                                                                                     ~                                                                                                                                     ~                                                                                                                                     ~                                                                                                                                     1,1All[?25hP+q436f\P+q6b75\P+q6b64\P+q6b72\P+q6b6c\P+q2332\P+q2334\P+q2569\P+q2a37\P+q6b31\[?12$p[?25l/3b3b/3b3b          [?25h[?25l/f8f8/f8f8          [?25h",,terminal_output
+15,110961,"TERMINAL",0,0,"[?25l::[?25h",,terminal_output
+16,111345,"TERMINAL",0,0,"w",,terminal_output
+17,111364,"TERMINAL",0,0,"q",,terminal_output
+18,111726,"TERMINAL",0,0,"\r[?25l[?2004l[>4;m"".git/MERGE_MSG"" 6L, 273B written\r\r\r\n[?1004l[?2004l[?1l>[?25h[>4;m[?1049l\rMerge made by the 'ort' strategy.\r\n",,terminal_output
+19,111773,"TERMINAL",0,0," jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_causal.sh       | 71 ++++++++++++++++++++++++++++++++++++\r\n .../berlin/coinrun/mila_submission/ablations/coinrun_dynamics_ffn_dim_ablation.sh    | 72 +++++++++++++++++++++++++++++++++++++\r\n jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_gt_actions.sh   | 71 ++++++++++++++++++++++++++++++++++++\r\n .../franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_no_cotraining.sh | 72 +++++++++++++++++++++++++++++++++++++\r\n .../franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_no_flash_attn.sh | 71 ++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu._sh       | 67 ++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_50k.sh    | 65 +++++++++++++++++++++++++++++++++\r\n .../breakout/noise_schedule_runs/causal/train_dyn_single_gpu_50k_no_noise_aug.sh     | 66 ++++++++++++++++++++++++++++++++++\r\n .../horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_gt_actions._sh   | 63 ++++++++++++++++++++++++++++++++\r\n .../breakout/noise_schedule_runs/causal/train_dyn_single_gpu_gt_actions_50k.sh       | 61 +++++++++++++++++++++++++++++++\r\n .../noise_schedule_runs/causal/train_dyn_single_gpu_gt_actions_smaller_lr_50k._sh    | 61 +++++++++++++++++++++++++++++++\r\n .../horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_no_noise_aug._sh | 68 +++++++++++++++++++++++++++++++++++\r\n .../breakout/noise_schedule_runs/causal/train_dyn_single_gpu_smaller_lr_50k._sh      | 65 +++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/breakout/noise_schedule_runs/full-prec/train_dyn_single_gpu.sh     |  2 --\r\n jobs/mihir/horeka/coinrun/ablations/train_dyn_default-grain-ablation.sh              | 81 +++++++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/coinrun/ablations/train_dyn_default-no-noise-main.sh               | 81 +++++++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/coinrun/ablations/train_dyn_default-no-noise.sh                    | 82 ++++++++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/coinrun/ablations/train_dyn_default-sqrt-ablation.sh               | 81 +++++++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/coinrun/default_runs/train_dyn_default.sh                          | 81 +++++++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh                          |  5 ++-\r\n jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh                    | 32 ++++++++++++++---\r\n jobs/mihir/horeka/minecraft/default_runs/train_lam_8_nodes.sbatch                    | 70 ++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/minecraft/default_runs/train_tokenizer_8_nodes.sbatch              | 71 ++++++++++++++++++++++++++++++++++++\r\n jobs/mihir/horeka/preprocessing/coinrun_chunked.sh                                   | 11 +++---\r\n jobs/mihir/horeka/preprocessing/coinrun_chunked_500m.sh                              | 21 +++++++++++\r\n jobs/mihir/horeka/preprocessing/doom_chunked.sh                                      | 23 ++++++++++++\r\n 26 files changed, 1499 insertions(+), 15 deletions(-)\r\n create mode 100644 jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_causal.sh\r\n create mode 100644 jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_ffn_dim_ablation.sh\r\n create mode 100644 jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_gt_actions.sh\r\n create mode 100644 jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_no_cotraining.sh\r\n create mode 100644 jobs/franz/berlin/coinrun/mila_submission/ablations/coinrun_dynamics_no_flash_attn.sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu._sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_50k.sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_50k_no_noise_aug.sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_gt_actions._sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_gt_actions_50k.sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_gt_actions_smaller_lr_50k._sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_no_noise_aug._sh\r\n create mode 100644 jobs/mihir/horeka/breakout/noise_schedule_runs/causal/train_dyn_single_gpu_smaller_lr_50k._sh\r\n create mode 100644 jobs/mihir/horeka/coinrun/ablations/train_dyn_default-grain-ablation.sh\r\n create mode 100644 jobs/mihir/horeka/coinrun/ablations/train_dyn_default-no-noise-main.sh\r\n create mode 100644 jobs/mihir/horeka/coinrun/ablations/train_dyn_default-no-noise.sh\r\n create mode 100644 jobs/mihir/horeka/coinrun/ablations/train_dyn_default-sqrt-ablation.sh\r\n create mode 100644 jobs/mihir/horeka/coinrun/default_runs/train_dyn_default.sh\r\n create mode 100644 jobs/mihir/horeka/minecraft/default_runs/train_lam_8_nodes.sbatch\r\n create mode 100644 jobs/mihir/horeka/minecraft/default_runs/train_tokenizer_8_nodes.sbatch\r\n create mode 100644 jobs/mihir/horeka/preprocessing/coinrun_chunked_500m.sh\r\n create mode 100644 jobs/mihir/horeka/preprocessing/doom_chunked.sh\r\n]0;franz.srambical@hai-login1:~/jafar/slurm",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-1b85b896-114f-4adc-b559-52f92a3a305a1762181559268-2025_11_03-15.52.54.585/source.csv

@@ -0,0 +1,5 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"/home/franz.srambical/jafar/slurm/jobs/franz/berlin/atari/data_upload/upload_to_hf.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=24:00:00\n#SBATCH --cpus-per-task=8\n#SBATCH --gres=gpu:1\n#SBATCH --output=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/atari/data_upload/%x_%j.log\n#SBATCH --error=/fast/project/HFMI_SynergyUnit/jafar_ws/logs/franz/atari/data_upload/%x_%j.log\n#SBATCH --job-name=upload_to_hf\n\nsource .venv/bin/activate\n\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/alien p-doom/atari-alien-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/amidar p-doom/atari-amidar-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/assault p-doom/atari-assault-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/asterix p-doom/atari-asterix-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/bank_heist p-doom/atari-bank_heist-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/battle_zone p-doom/atari-battle_zone-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/boxing p-doom/atari-boxing-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/breakout p-doom/atari-breakout-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/chopper_command p-doom/atari-chopper_command-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/crazy_climber p-doom/atari-crazy_climber-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/demon_attack p-doom/atari-demon_attack-dataset --repo-type dataset\npython slurm/utils/mihir/upload_hf_dataset.py /fast/project/HFMI_SynergyUnit/jafar_ws/data/atari/pong p-doom/atari-pong-dataset --repo-type dataset",shellscript,tab
+2,399,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:52:54 PM [info] Activating crowd-code\n3:52:54 PM [info] Recording started\n3:52:54 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,602,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"3:52:54 PM [info] Git repository found\n3:52:54 PM [info] Git provider initialized successfully\n3:52:54 PM [info] Initial git state: [object Object]\n",Log,content
+4,765849,"/home/franz.srambical/jafar/slurm/jobs/franz/berlin/atari/data_upload/upload_to_hf.sh",0,0,"",shellscript,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-1c973de0-6be0-4d96-a6b9-3da6c2c9d2721756538623658-2025_08_30-08.23.47.981/source.csv

@@ -0,0 +1,5 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,48,"tasks",0,0,"",Log,tab
+3,96,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+4,1503,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:23:47 AM [info] Activating crowd-code\n8:23:47 AM [info] Recording started\n8:23:47 AM [info] Initializing git provider using file system watchers...\n8:23:47 AM [info] No workspace folder found\n",Log,content
+5,2029,"extension-output-pdoom-org.crowd-code-#1-crowd-code",194,0,"8:23:49 AM [info] Retrying git provider initialization...\n8:23:49 AM [info] No workspace folder found\n",Log,content

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-20194a13-4640-4c1b-9e26-6b88dd86963a1766586898739-2025_12_24-15.35.09.94/source.csv

@@ -0,0 +1,14 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"scripts/run-sft-torchrun.sh",0,0,"#!/bin/bash\n#\n# Ray-free SFT Training Script\n#\nexport PYTHONUNBUFFERED=1\nexport PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True\n# FIXME(f.srambical): this is hardcoded for now\nGPUS_PER_NODE=${SLURM_GPUS_ON_NODE}\nNUM_NODES=${SLURM_JOB_NUM_NODES}\nNODE_RANK=${SLURM_NODEID}\nMASTER_ADDR=${MASTER_ADDR:-$(scontrol show hostnames $SLURM_JOB_NODELIST | head -n 1)}\n\nNVLINK_COUNT=$(nvidia-smi | grep -o ""NVLink"" | wc -l)\nif [ ""$NVLINK_COUNT"" -gt 0 ]; then\n    HAS_NVLINK=1\nelse\n    HAS_NVLINK=0\nfi\necho ""HAS_NVLINK: $HAS_NVLINK (detected $NVLINK_COUNT NVLink references)""\n\nexport NCCL_DEBUG=INFO\nexport TORCH_DISTRIBUTED_DEBUG=INFO\n\nSCRIPT_DIR=""$(cd -- ""$(dirname -- ""${BASH_SOURCE[0]}"")"" &>/dev/null && pwd)""\n\nRUN_ID=${RUN_ID:-""run_$(date +%Y%m%d_%H%M%S)""}\nLOAD_PATH=""/fast/project/HFMI_SynergyUnit/tab_model/huggingface/Qwen3-0.6B""\nSAVE_PATH=""/fast/project/HFMI_SynergyUnit/tab_model/huggingface/shared_data/${RUN_ID}/checkpoints""\n\nCKPT_ARGS=(\n   --hf-checkpoint /fast/project/HFMI_SynergyUnit/tab_model/huggingface/Qwen3-0.6B\n   --load ${LOAD_PATH}\n   --ref-load /fast/project/HFMI_SynergyUnit/tab_model/huggingface/Qwen3-0.6B\n   --save ${SAVE_PATH}\n   --save-interval 1000\n)\n\nSFT_ARGS=(\n   --rollout-function-path miles.rollout.sft_rollout.generate_rollout\n   --prompt-data /fast/project/HFMI_SynergyUnit/tab_model/huggingface/nemo_hf_part_jsonl_4k_tokens.jsonl\n   --val-prompt-data /fast/project/HFMI_SynergyUnit/tab_model/huggingface/nemo_hf_part_jsonl_4k_tokens_validation.jsonl\n   --val-interval 1000\n   --val-steps 100\n   --input-key messages\n   --apply-chat-template\n   --rollout-shuffle\n   --num-rollout 10000\n   --rollout-batch-size 16\n   --global-batch-size 16\n\n   --loss-type sft_loss\n   --calculate-per-token-loss\n   --disable-compute-advantages-and-returns\n)\n\nLORA_ARGS=(\n    --use-lora\n    --lora-rank 8\n    --lora-alpha 16\n    --lora-dropout 0.0\n    --lora-target-modules q_proj v_proj\n)\n\nOPTIMIZER_ARGS=(\n   --optimizer adam\n   --lr 1e-4\n   --lr-decay-style WSD\n   --lr-wsd-decay-style linear\n   --lr-warmup-iters 500\n   --lr-decay-iters 10000\n   --lr-wsd-decay-iters 2000\n   --weight-decay 0.1\n   --adam-beta1 0.9\n   --adam-beta2 0.98\n)\n\nWANDB_ARGS=(\n   --use-wandb\n   --wandb-project crowd-pilot-miles\n   --wandb-team instant-uv\n   --wandb-group qwen3-0.6b-sft-torchrun\n)\n\nTRAIN_BACKEND_ARGS=(\n   --train-backend fsdp\n   --update-weight-buffer-size 536870912\n   --attn-implementation flash_attention_3\n)\n\nPERF_ARGS=(\n   --use-dynamic-batch-size\n   --max-tokens-per-gpu 9216\n)\n\nMISC_ARGS=(\n   --rollout-max-context-len 8192\n   --rollout-max-prompt-len 8000\n   --rollout-max-response-len 8192\n   --dump-details /fast/project/HFMI_SynergyUnit/tab_model/huggingface/shared_data/qwen3-600M-fsdp-1116-noref/dump_details\n)\n\ntorchrun \\n    --nproc_per_node=${GPUS_PER_NODE} \\n    --nnodes=${NUM_NODES} \\n    --node_rank=${NODE_RANK} \\n    --master_addr=${MASTER_ADDR} \\n    --master_port=${MASTER_PORT:-29500} \\n    train_sft.py \\n    ${CKPT_ARGS[@]} \\n    ${SFT_ARGS[@]} \\n    ${LORA_ARGS[@]} \\n    ${OPTIMIZER_ARGS[@]} \\n    ${WANDB_ARGS[@]} \\n    ${TRAIN_BACKEND_ARGS[@]} \\n    ${PERF_ARGS[@]} \\n    ${MISC_ARGS[@]}\n",shellscript,tab
+2,128,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:35:09 PM [info] Activating crowd-code\n3:35:09 PM [info] Recording started\n3:35:09 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,175,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"3:35:09 PM [info] Git repository found\n3:35:09 PM [info] Git provider initialized successfully\n3:35:09 PM [info] Initial git state: [object Object]\n",Log,content
+4,538023,"TERMINAL",0,0,"",,terminal_focus
+5,538025,"scripts/run-sft-torchrun.sh",0,0,"",shellscript,tab
+6,538730,"TERMINAL",0,0,"source /home/franz.srambical/miles/.venv/bin/activate",,terminal_command
+7,538739,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~/miles",,terminal_output
+8,539525,"TERMINAL",0,0,"squeue",,terminal_command
+9,539548,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  38543   xiao.liu interacti     1    64 PD 2025-12-24T10:38:15 2025-12-24T16:01:06     0:00   23:59:00 (Resources)\r\n  38561 alfred.ngu interacti     1     4 PD 2025-12-24T15:38:02                 N/A     0:00    2:00:00 (Priority)\r\n  38556 mihir.maha interacti     1     2  R 2025-12-24T13:44:18 2025-12-24T13:44:22  1:59:46    2:00:00 hai006\r\n  38526   xiao.liu interacti     1   128  R 2025-12-24T01:52:12 2025-12-24T01:52:12 13:51:56   23:59:00 hai002\r\n  38525   xiao.liu interacti     1   128  R 2025-12-24T01:51:21 2025-12-24T01:51:21 13:52:47   23:59:00 hai003\r\n  38518 franz.sram interacti     1   200  R 2025-12-23T21:54:04 2025-12-23T21:54:04 17:50:04 1-00:00:00 hai007\r\n  38505   xiao.liu interacti     1   128  R 2025-12-23T16:02:06 2025-12-23T16:02:06 23:42:02   23:59:00 hai006\r\n  38527 nishant.ku  standard     3   624  R 2025-12-24T02:30:11 2025-12-24T10:38:14  5:05:54 1-00:00:00 hai[001,004-005]\r\n  38528   xiao.liu  standard     1   128  R 2025-12-24T03:28:17 2025-12-24T03:28:18 12:15:50   23:59:00 hai008\r\n]0;franz.srambical@hai-login2:~/miles",,terminal_output
+10,544358,"TERMINAL",0,0,"scancel --me",,terminal_command
+11,544386,"TERMINAL",0,0,"]633;C]0;franz.srambical@hai-login2:~/miles",,terminal_output
+12,571012,"TERMINAL",0,0,"squeue",,terminal_command
+13,571017,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n  38518 franz.sram interacti     1   200 CG 2025-12-23T21:54:04 2025-12-23T21:54:04 17:50:09 1-00:00:00 hai007\r\n  38543   xiao.liu interacti     1    64 PD 2025-12-24T10:38:15 2025-12-24T15:44:40     0:00   23:59:00 (Resources)\r\n  38561 alfred.ngu interacti     1     4 PD 2025-12-24T15:38:02                 N/A     0:00    2:00:00 (Priority)\r\n  38556 mihir.maha interacti     1     2  R 2025-12-24T13:44:18 2025-12-24T13:44:22  2:00:18    2:00:00 hai006\r\n  38526   xiao.liu interacti     1   128  R 2025-12-24T01:52:12 2025-12-24T01:52:12 13:52:28   23:59:00 hai002\r\n  38525   xiao.liu interacti     1   128  R 2025-12-24T01:51:21 2025-12-24T01:51:21 13:53:19   23:59:00 hai003\r\n  38505   xiao.liu interacti     1   128  R 2025-12-23T16:02:06 2025-12-23T16:02:06 23:42:34   23:59:00 hai006\r\n  38527 nishant.ku  standard     3   624  R 2025-12-24T02:30:11 2025-12-24T10:38:14  5:06:26 1-00:00:00 hai[001,004-005]\r\n  38528   xiao.liu  standard     1   128  R 2025-12-24T03:28:17 2025-12-24T03:28:18 12:16:22   23:59:00 hai008\r\n]0;franz.srambical@hai-login2:~/miles",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-4d667d63-df3b-4c2f-9829-cbc24fb4ff5d1767609520115-2026_01_05-11.38.46.560/source.csv

@@ -0,0 +1,6 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize_16k_glm.sh",0,0,"./target/release/crowd-pilot-serialize \\n--csv-root=""/fast/project/HFMI_SynergyUnit/tab_model/data/hf_part_csv/"" \\n--output-dir=""/fast/project/HFMI_SynergyUnit/tab_model/data/glm/miles_hf_part_jsonl_16k_tokens/"" \\n--max-tokens-per-conversation 16384 \\n--tokenizer=""zai-org/GLM-4.5-Air""\n",shellscript,tab
+2,214,"tasks",0,0,"",Log,tab
+3,215,"/home/franz.srambical/slurm/dev/franz/berlin/crowd-pilot/crowd_pilot_serializer/serialize_16k_glm.sh",0,0,"",shellscript,tab
+4,222,"TERMINAL",0,0,"",,terminal_focus
+5,235,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"11:38:46 AM [info] Activating crowd-code\n11:38:46 AM [info] Recording started\n11:38:46 AM [info] Initializing git provider using file system watchers...\n11:38:46 AM [info] Git repository found\n11:38:46 AM [info] Git provider initialized successfully\n11:38:46 AM [info] Initial git state: [object Object]\n",Log,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-59175e55-ecae-446f-be12-8861032d4f481751613426266-2025_07_04-09.17.44.620/source.csv

@@ -0,0 +1,16 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,5,"tests/test_checkpointer.py",0,0,"import unittest\nimport tempfile\nimport os\nimport jax\nimport jax.numpy as jnp\nfrom flax.training import orbax_utils\nfrom orbax.checkpoint import PyTreeCheckpointer\nfrom pathlib import Path\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom flax.training.train_state import TrainState\nimport optax\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\n\nclass DistributedCheckpointerTest(unittest.TestCase):\n    def setUp(self):\n        super().setUp()\n        self._temp_dir_manager = tempfile.TemporaryDirectory()\n        self.checkpoint_dir = Path(self._temp_dir_manager.name)\n        self.addCleanup(self._temp_dir_manager.cleanup)\n\n        # FIXME (f.srambical): If the tests pass, we should use the default model config instead\n        self.model_kwargs = dict(\n            in_dim=3,\n            model_dim=8,\n            latent_dim=4,\n            num_latents=16,\n            patch_size=2,\n            num_blocks=1,\n            num_heads=1,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.image_shape = (8, 8, 3)\n        self.seq_len = 2\n        self.batch_size = 2\n        self.seed = 0\n\n    def test_distributed_checkpointing(self):\n        jax.distributed.initialize()\n        num_devices = jax.device_count()\n        self.assertGreater(num_devices, 0)\n\n        model = TokenizerVQVAE(**self.model_kwargs)\n        rng = jax.random.PRNGKey(self.seed)\n        dummy_inputs = dict(\n            videos=jnp.zeros((self.batch_size, self.seq_len, *self.image_shape), dtype=jnp.float32)\n        )\n        params = model.init(rng, dummy_inputs)\n\n        tx = optax.adam(1e-3)\n        state = TrainState.create(apply_fn=model.apply, params=params, tx=tx)\n\n        device_mesh_arr = create_device_mesh((num_devices,))\n        mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n        replicated_sharding = NamedSharding(mesh, PartitionSpec())\n        state = jax.device_put(state, replicated_sharding)\n\n        ckpt = {""model"": state}\n        orbax_checkpointer = PyTreeCheckpointer()\n        save_args = orbax_utils.save_args_from_target(ckpt)\n        ckpt_path = str(self.checkpoint_dir / ""test_ckpt"")\n        orbax_checkpointer.save(ckpt_path, ckpt, save_args=save_args)\n        self.assertTrue(os.path.exists(ckpt_path))\n\n        restore_target = {""model"": state}\n        restore_args = orbax_utils.restore_args_from_target(restore_target)\n        restored = orbax_checkpointer.restore(ckpt_path, item=restore_target, restore_args=restore_args)\n        # Compare parameters recursively, handling nested structure\n        def compare_params(original, restored):\n            if isinstance(original, dict):\n                for k in original.keys():\n                    compare_params(original[k], restored[k])\n            else:\n                self.assertTrue(jnp.allclose(original, restored))\n        \n        compare_params(state.params, restored[""model""].params)\n\nif __name__ == ""__main__"":\n    unittest.main()\n",python,tab
+2,568,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:17:43 AM [info] Activating crowd-code\n9:17:44 AM [info] Recording started\n9:17:44 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1547,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"9:17:45 AM [info] Git repository found\n9:17:45 AM [info] Git provider initialized successfully\n9:17:45 AM [info] Initial git state: [object Object]\n",Log,content
+4,1648,"tests/test_checkpointer.py",0,0,"",python,tab
+5,3081,"TERMINAL",0,0,"/usr/bin/python3 /ictstr01/home/aih/franz.srambical/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /ictstr01/home/aih/franz.srambical/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt",,terminal_command
+6,3095,"TERMINAL",0,0,"]633;E;/usr/bin/python3 /ictstr01/home/aih/franz.srambical/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /ictstr01/home/aih/franz.srambical/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt;b730ba2f-be2c-4b6f-9f0d-d578d409e7ab]633;C]0;franz.srambical@hpc-submit01:/ictstr01/home/aih/franz.srambical/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash]633;D;0]633;P;Cwd=/ictstr01/home/aih/franz.srambical/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash",,terminal_output
+7,13336,"TERMINAL",0,0,"salloc --reservation=haicu_stefan      -p gpu_p    --time=05:00:00      --job-name=interactive_bash  --qos=gpu_normal --gres=gpu:4 -w supergpu16 --cpus-per-task=8",,terminal_command
+8,13419,"TERMINAL",0,0,"]633;E;salloc --reservation=haicu_stefan -p gpu_p --time=05:00:00 --job-name=interactive_bash --qos=gpu_normal --gres=gpu:4 -w supergpu16 --cpus-per-task=8;a8707c05-ae9b-4a50-91c9-fa9c06501dad]633;Csalloc: Required node not available (down, drained or reserved)\r\nsalloc: Pending job allocation 26666565\r\nsalloc: job 26666565 queued and waiting for resources\r\n",,terminal_output
+9,79296,"TERMINAL",0,0,"^Csalloc: Job allocation 26666565 has been revoked.\r\nsalloc: Job aborted due to signal\r\n",,terminal_output
+10,82221,"TERMINAL",0,0,"sinfo supergpu16",,terminal_command
+11,82252,"TERMINAL",0,0,"\r\n[?2004l\r]633;E;sinfo supergpu16;a8707c05-ae9b-4a50-91c9-fa9c06501dad]633;CPARTITION         AVAIL  TIMELIMIT  NODES  STATE NODELIST\r\ninteractive_cpu_p    up   12:00:00      1  down* cpusrv54\r\ninteractive_cpu_p    up   12:00:00      1    mix cpusrv75\r\ninteractive_cpu_p    up   12:00:00      4   idle cpusrv[51-53,55]\r\ncpu_p                up 3-00:00:00      1  down* supercpu02\r\ncpu_p                up 3-00:00:00      2  drain cpusrv[57,74]\r\ncpu_p                up 3-00:00:00     84    mix cpusrv[02,05-27,31-33,35-39,41-47,49-50,56,58,61-65,72-73,75,77,79,82-83,89,92,94,96-104,106-108,110,112-113,115-117,119,121-122,124-127],supercpu01\r\ncpu_p                up 3-00:00:00     19  alloc cpusrv[28,59-60,78,80-81,84-88,90-91,93,109,111,114,118,123]\r\ncpu_p                up 3-00:00:00      8   idle cpusrv[30,40,48,69-71,95,105]\r\ninteractive_gpu_p    up   12:00:00      1  inval gpusrv25\r\ninteractive_gpu_p    up   12:00:00      3   idle gpusrv[22-24]\r\ngpu_p                up 2-00:00:00      3   mix- supergpu[14,17,19]\r\ngpu_p                up 2-00:00:00      2  down* gpusrv34,supergpu07\r\ngpu_p                up 2-00:00:00      1  drain supergpu16\r\ngpu_p                up 2-00:00:00      2   resv supergpu[05,18]\r\ngpu_p                up 2-00:00:00     52    mix gpusrv[11-12,15,18,26-33,38-40,42-46,50-55,57-77],supergpu[02-03,08-09,15]\r\ngpu_p                up 2-00:00:00     11   idle gpusrv[09-10,13-14,16-17,35,41,47-49]\r\ncemp_gpu_p           up 5-00:00:00      1  down* supercpu02\r\ncemp_gpu_p           up 5-00:00:00      3    mix supercpu01,supergpu[06,10]\r\ncemp_gpu_p           up 5-00:00:00      3   idle supergpu[11-13]\r\nbcf_p                up 14-00:00:0      1    mix cpusrv29\r\nbcf_p                up 14-00:00:0      1   idle cpusrv128\r\n]0;franz.srambical@hpc-submit01:/lustre/groups/haicu/workspace/franz.srambical/jafar]633;D;0",,terminal_output
+12,144962,"TERMINAL",0,0,"scontrol show node supergpu16~",,terminal_command
+13,144987,"TERMINAL",0,0,"]633;E;scontrol show node supergpu16~;a8707c05-ae9b-4a50-91c9-fa9c06501dad]633;C",,terminal_output
+14,147337,"TERMINAL",0,0,"scontrol show node supergpu16",,terminal_command
+15,147379,"TERMINAL",0,0,"\r\n[?2004l\r]633;E;scontrol show node supergpu16;a8707c05-ae9b-4a50-91c9-fa9c06501dad]633;C",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-65bd7149-5424-43f9-8618-556923c89f3e1762712224460-2025_11_09-19.17.10.464/source.csv

@@ -0,0 +1,4 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,4,"crowd-pilot-extension/src/extension.ts",0,0,"import * as vscode from 'vscode';\nimport * as http from 'http';\nimport { Buffer } from 'buffer';\n\nconst HOSTNAME = 'hai005';\nconst PORT = 30000;\n\nexport function activate(context: vscode.ExtensionContext) {\n\n\tconsole.log('[crowd-pilot] Extension activated');\n\n\t// Configure terminal to allow tab keybinding to work\n\t(async () => {\n\t\tconst config = vscode.workspace.getConfiguration('terminal.integrated');\n\t\tconst commandsToSkipShell = config.get<string[]>('commandsToSkipShell', []);\n\t\tlet updated = false;\n\t\tif (!commandsToSkipShell.includes('crowd-pilot.modelRun')) {\n\t\t\tcommandsToSkipShell.push('crowd-pilot.modelRun');\n\t\t\tupdated = true;\n\t\t}\n\t\tif (!commandsToSkipShell.includes('crowd-pilot.hideUi')) {\n\t\t\tcommandsToSkipShell.push('crowd-pilot.hideUi');\n\t\t\tupdated = true;\n\t\t}\n\t\tif (updated) {\n\t\t\tawait config.update('commandsToSkipShell', commandsToSkipShell, vscode.ConfigurationTarget.Global);\n\t\t}\n\t})().catch((err) => console.error('[crowd-pilot] Startup initialization error:', err));\n\n\tconst hideUi = vscode.commands.registerCommand('crowd-pilot.hideUi', () => {\n\t\thidePreviewUI();\n\t});\n\n\tconst modelRun = vscode.commands.registerCommand('crowd-pilot.modelRun', async () => {\n\t\tconst editor = vscode.window.activeTextEditor;\n\t\tif (!editor) {\n\t\t\treturn;\n\t\t}\n\t\ttry {\n\t\t\tconst plan = await requestModelActions(editor);\n\n\t\t\tif (!previewVisible) {\n\t\t\t\tshowPreviewUI(plan);\n\t\t\t\treturn;\n\t\t\t}\n\n\t\t\tconst runPlan = currentPlan ?? plan;\n\t\t\thidePreviewUI();\n\t\t\tawait executePlan(runPlan);\n\t\t\tvscode.window.showInformationMessage('All actions emitted');\n\t\t} catch (err) {\n\t\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\tvscode.window.showErrorMessage(`Model run failed: ${errorMessage}`);\n\t\t}\n\t});\n\n\tconst sglangTest = vscode.commands.registerCommand('crowd-pilot.sglangTest', async () => {\n\t\ttry {\n\t\t\tawait callSGLangChat();\n\t\t} catch (err) {\n\t\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\tvscode.window.showErrorMessage(`SGLang test failed: ${errorMessage}`);\n\t\t}\n\t});\n\n\tcontext.subscriptions.push(hideUi, sglangTest, modelRun);\n}\n\nexport function deactivate() {}\n\n// -------------------- Plan Types & Execution --------------------\ntype PlannedAction =\n| { kind: 'showTextDocument' }\n| { kind: 'setSelections', selections: Array<{ start: [number, number], end: [number, number] }> }\n| { kind: 'editInsert', position: [number, number], text: string }\n| { kind: 'terminalShow' }\n| { kind: 'terminalSendText', text: string };\n\nlet currentPlan: PlannedAction[] | undefined;\n\nasync function executePlan(plan: PlannedAction[]): Promise<void> {\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\tconst doc = editor.document;\n\tconst term = vscode.window.terminals[0] ?? vscode.window.createTerminal('Test');\n\tfor (const action of plan) {\n\t\tif (action.kind === 'showTextDocument') {\n\t\t\tawait vscode.window.showTextDocument(doc);\n\t\t\tcontinue;\n\t\t}\n\t\tif (action.kind === 'setSelections') {\n\t\t\teditor.selections = action.selections.map(s => new vscode.Selection(\n\t\t\t\tnew vscode.Position(s.start[0], s.start[1]),\n\t\t\t\tnew vscode.Position(s.end[0], s.end[1])\n\t\t\t));\n\t\t\tcontinue;\n\t\t}\n\t\tif (action.kind === 'editInsert') {\n\t\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.insert(new vscode.Position(action.position[0], action.position[1]), action.text));\n\t\t\tcontinue;\n\t\t}\n\t\tif (action.kind === 'terminalShow') {\n\t\t\tterm.show();\n\t\t\tcontinue;\n\t\t}\n\t\tif (action.kind === 'terminalSendText') {\n\t\t\tterm.sendText(action.text);\n\t\t\tcontinue;\n\t\t}\n\t}\n}\n\n// -------------------- UI State & Helpers --------------------\nconst UI_CONTEXT_KEY = 'crowdPilot.uiVisible';\nlet previewVisible = false;\nlet previewQuickPick: vscode.QuickPick<(vscode.QuickPickItem & { index: number })> | undefined;\n\nfunction showPreviewUI(plan: PlannedAction[]): void {\n\tconst items: (vscode.QuickPickItem & { index: number })[] = plan.map((action, index) => {\n\t\tswitch (action.kind) {\n\t\t\tcase 'showTextDocument':\n\t\t\t\treturn { index, label: '$(file) Focus active text document' };\n\t\t\tcase 'setSelections':\n\t\t\t\t{\n\t\t\t\t\tconst cursors = action.selections.map(s => `(${s.start[0]}, ${s.start[1]})`).join(', ');\n\t\t\t\t\treturn { index, label: `$(cursor) Move cursor to ${cursors}` };\n\t\t\t\t}\n\t\t\tcase 'editInsert':\n\t\t\t\treturn { index, label: `$(pencil) Insert ""${action.text.replace(/\n/g, '\\n')}"" at (${action.position[0]}, ${action.position[1]})` };\n\t\t\tcase 'terminalShow':\n\t\t\t\treturn { index, label: '$(terminal) Focus terminal' };\n\t\t\tcase 'terminalSendText':\n\t\t\t\treturn { index, label: `$(terminal) Run ""${action.text}"" in terminal` };\n\t\t}\n\t});\n    if (!previewQuickPick) {\n        previewQuickPick = vscode.window.createQuickPick<(vscode.QuickPickItem & { index: number })>();\n\t\tpreviewQuickPick.title = 'crowd-pilot: preview';\n\t\tpreviewQuickPick.matchOnDetail = true;\n\t\tpreviewQuickPick.ignoreFocusOut = true;\n\t\tpreviewQuickPick.canSelectMany = false;\n        previewQuickPick.onDidAccept(async () => {\n            const qp = previewQuickPick!;\n            const selected = qp.selectedItems?.[0];\n            qp.hide();\n            if (selected) {\n                await executePlan([plan[selected.index]]);\n                vscode.window.showInformationMessage('Action executed');\n            }\n        });\n\t\tpreviewQuickPick.onDidHide(() => {\n\t\t\tpreviewVisible = false;\n\t\t\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n\t\t\ttry { previewQuickPick?.dispose(); } catch {}\n\t\t\tpreviewQuickPick = undefined;\n\t\t});\n\t}\n\tpreviewQuickPick.items = items;\n\tpreviewQuickPick.placeholder = 'Press Tab to run all, Enter for selected, or Esc to hide';\n\tpreviewQuickPick.show();\n\tpreviewVisible = true;\n\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, true);\n\tcurrentPlan = plan;\n}\n\nfunction hidePreviewUI(): void {\n\tif (previewQuickPick) {\n\t\ttry { previewQuickPick.hide(); } catch {}\n\t\treturn;\n\t}\n\tpreviewVisible = false;\n\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n}\n\n// -------------------- SGLang Client (simple test) --------------------\nasync function callSGLangChat(): Promise<void> {\n\tconst requestBody = {\n\t\tmodel: 'qwen/qwen2.5-0.5b-instruct',\n\t\tmessages: [\n\t\t\t{ role: 'user', content: 'What is the capital of France?' }\n\t\t]\n\t};\n\tconst postData = JSON.stringify(requestBody);\n\n\tconst options = {\n\t\thostname: HOSTNAME,\n\t\tport: PORT,\n\t\tpath: '/v1/chat/completions',\n\t\tmethod: 'POST',\n\t\theaders: {\n\t\t\t'Content-Type': 'application/json',\n\t\t\t'Content-Length': Buffer.byteLength(postData)\n\t\t}\n\t};\n\n\ttry {\n\t\tconst json = await new Promise<any>((resolve, reject) => {\n\t\t\tconst req = http.request(options, (res: http.IncomingMessage) => {\n\t\t\t\tlet data = '';\n\t\t\t\tres.on('data', (chunk: Buffer) => {\n\t\t\t\t\tdata += chunk.toString();\n\t\t\t\t});\n\t\t\t\tres.on('end', () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tresolve(JSON.parse(data));\n\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\treject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t});\n\n\t\t\treq.on('error', (err: Error) => {\n\t\t\t\treject(err);\n\t\t\t});\n\n\t\t\treq.write(postData);\n\t\t\treq.end();\n\t\t});\n\n\t\tvscode.window.showInformationMessage(`SGLang response: ${JSON.stringify(json, null, 2)}`);\n\t} catch (err) {\n\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\tvscode.window.showErrorMessage(`SGLang request failed: ${errorMessage}`);\n\t}\n}\n\n// -------------------- Model-planned Actions --------------------\nasync function requestModelActions(editor: vscode.TextEditor): Promise<PlannedAction[]> {\n\tconst schemaDescription = [\n\t\t'Role: You suggest the next VS Code editor/terminal action to progress the current task.',\n\t\t'Output ONLY a JSON array (no prose, no code fences). Length exactly 1.',\n\t\t'Coordinates are zero-based [line, column].',\n\t\t'Allowed actions (JSON schema-like):',\n\t\t'{ kind: ""showTextDocument"" }',\n\t\t'{ kind: ""setSelections"", selections: Array<{ start: [number, number], end: [number, number] }> }',\n\t\t'{ kind: ""editInsert"", position: [number, number], text: string }',\n\t\t'{ kind: ""terminalShow"" }',\n\t\t'{ kind: ""terminalSendText"", text: string }',\n\t\t'Guidelines:',\n\t\t'- If you you insert text, insert until the logical end of the current statement or block.',\n\t\t'- When inserting text, make sure to not repeat existing text (except when replacing existing text).',\n\t\t'- Use double-quoted JSON strings.'\n\t].join('\n');\n\n\tconst doc = editor.document;\n\tconst cursor = editor.selection.active;\n\tconst contextRange = new vscode.Range(new vscode.Position(0, 0), cursor);\n\tconst contextCode = doc.getText(contextRange);\n\tconst maxContextChars = 20000;\n\tconst allLines = contextCode.split(/\r?\n/);\n\tlet startLineIndex = 0;\n\tlet visibleLines = allLines;\n\tif (contextCode.length > maxContextChars) {\n\t\tlet acc = 0;\n\t\tlet idx = allLines.length;\n\t\twhile (idx > 0 && acc <= maxContextChars) {\n\t\t\tidx--;\n\t\t\tacc += allLines[idx].length + 1;\n\t\t}\n\t\tstartLineIndex = idx;\n\t\tvisibleLines = allLines.slice(idx);\n\t}\n\tconst numberedContext = visibleLines.map((line, i) => `${startLineIndex + i}: ${line}`).join('\n');\n\n\tconst tabbingPrompt = [\n\t\t'Your role: Propose the single next action according to the schema to help the developer progress.',\n\t\t'',\n\t\t'Available context:',\n\t\t`- File: ${doc.fileName}`,\n\t\t`- Language: ${doc.languageId}`,\n\t\t`- Cursor: (${cursor.line}, ${cursor.character})`,\n\t\t'',\n\t\t'Current file content up to the cursor (zero-based line numbers):',\n\t\t'```',\n\t\tnumberedContext,\n\t\t'```',\n\t\t'',\n\t\t'Respond with ONLY a JSON array containing exactly one action.'\n\t].join('\n');\n\n\tconst requestBody = {\n\t\tmodel: 'qwen/qwen2.5-0.5b-instruct',\n\t\tmessages: [\n\t\t\t{ role: 'system', content: schemaDescription },\n\t\t\t{ role: 'user', content: tabbingPrompt }\n\t\t]\n\t};\n\n\tconst postData = JSON.stringify(requestBody);\n\tconst options = {\n\t\thostname: HOSTNAME,\n\t\tport: PORT,\n\t\tpath: '/v1/chat/completions',\n\t\tmethod: 'POST',\n\t\theaders: {\n\t\t\t'Content-Type': 'application/json',\n\t\t\t'Content-Length': Buffer.byteLength(postData)\n\t\t}\n\t};\n\n\tconst json = await new Promise<any>((resolve, reject) => {\n\t\tconst req = http.request(options, (res: http.IncomingMessage) => {\n\t\t\tlet data = '';\n\t\t\tres.on('data', (chunk: Buffer) => { data += chunk.toString(); });\n\t\t\tres.on('end', () => {\n\t\t\t\ttry {\n\t\t\t\t\tresolve(JSON.parse(data));\n\t\t\t\t} catch (err) {\n\t\t\t\t\treject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n\t\t\t\t}\n\t\t\t});\n\t\t});\n\t\treq.on('error', (err: Error) => reject(err));\n\t\treq.write(postData);\n\t\treq.end();\n\t});\n\n\tconst content = extractChatContent(json);\n\tif (typeof content !== 'string' || content.trim().length === 0) {\n\t\tthrow new Error('Empty model content');\n\t}\n\tconst actions = parsePlannedActions(content);\n\tif (actions.length === 0) {\n\t\tthrow new Error('No valid actions parsed from model output');\n\t}\n\treturn actions;\n}\n\nfunction extractChatContent(json: any): string | undefined {\n\ttry {\n\t\tif (json && Array.isArray(json.choices) && json.choices[0]) {\n\t\t\tconst choice = json.choices[0];\n\t\t\tif (choice.message && typeof choice.message.content === 'string') {\n\t\t\t\treturn choice.message.content;\n\t\t\t}\n\t\t\tif (typeof choice.text === 'string') {\n\t\t\t\treturn choice.text;\n\t\t\t}\n\t\t}\n\t\treturn undefined;\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction parsePlannedActions(raw: string): PlannedAction[] {\n\tlet text = raw.trim();\n\ttext = text.replace(/^```(?:json)?\s*/i, '').replace(/```\s*$/i, '').trim();\n\tconst arrayMatch = text.match(/\[[\s\S]*\]/);\n\tconst jsonText = arrayMatch ? arrayMatch[0] : text;\n\tlet parsed: unknown;\n\ttry {\n\t\tparsed = JSON.parse(jsonText);\n\t} catch (err) {\n\t\treturn [];\n\t}\n\tif (!Array.isArray(parsed)) { return []; }\n\tconst result: PlannedAction[] = [];\n\tfor (const item of parsed) {\n\t\tif (!item || typeof item !== 'object' || typeof (item as any).kind !== 'string') { continue; }\n\t\tswitch ((item as any).kind) {\n\t\t\tcase 'showTextDocument':\n\t\t\t\tresult.push({ kind: 'showTextDocument' });\n\t\t\t\tbreak;\n\t\t\tcase 'setSelections': {\n\t\t\t\tconst selections = Array.isArray((item as any).selections) ? (item as any).selections : [];\n\t\t\t\tconst norm = selections.map((s: any) => ({\n\t\t\t\t\tstart: Array.isArray(s?.start) && s.start.length === 2 ? [Number(s.start[0]) || 0, Number(s.start[1]) || 0] as [number, number] : [0, 0] as [number, number],\n\t\t\t\t\tend: Array.isArray(s?.end) && s.end.length === 2 ? [Number(s.end[0]) || 0, Number(s.end[1]) || 0] as [number, number] : [0, 0] as [number, number]\n\t\t\t\t}));\n\t\t\t\tresult.push({ kind: 'setSelections', selections: norm });\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase 'editInsert': {\n\t\t\t\tconst pos = Array.isArray((item as any).position) && (item as any).position.length === 2 ? [Number((item as any).position[0]) || 0, Number((item as any).position[1]) || 0] as [number, number] : [0, 0] as [number, number];\n\t\t\t\tconst text = typeof (item as any).text === 'string' ? (item as any).text : '';\n\t\t\t\tresult.push({ kind: 'editInsert', position: pos, text });\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase 'terminalShow':\n\t\t\t\tresult.push({ kind: 'terminalShow' });\n\t\t\t\tbreak;\n\t\t\tcase 'terminalSendText': {\n\t\t\t\tconst text = typeof (item as any).text === 'string' ? (item as any).text : '';\n\t\t\t\tresult.push({ kind: 'terminalSendText', text });\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tdefault:\n\t\t\t\tbreak;\n\t\t}\n\t}\n\treturn result;\n}\n",typescript,tab
+2,199,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:17:10 PM [info] Activating crowd-code\n7:17:10 PM [info] Recording started\n7:17:10 PM [info] Initializing git provider using file system watchers...\n7:17:10 PM [info] Git repository found\n7:17:10 PM [info] Git provider initialized successfully\n7:17:10 PM [info] Initial git state: [object Object]\n",Log,tab
+3,2876,"crowd-pilot-extension/src/extension.ts",0,0,"",typescript,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-667f558c-c85b-45c3-9621-f920186c4d561764772496749-2025_12_03-15.35.04.989/source.csv

@@ -0,0 +1,6 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"src/extension.ts",0,0,"import * as vscode from 'vscode';\nimport * as http from 'http';\nimport { Buffer } from 'buffer';\n\nconst HOSTNAME = 'hai001';\nconst PORT = 30000;\n\nexport function activate(context: vscode.ExtensionContext) {\n\n\tconsole.log('[crowd-pilot] Extension activated');\n\n\t// Configure terminal to allow tab keybinding to work\n\t(async () => {\n\t\tconst config = vscode.workspace.getConfiguration('terminal.integrated');\n\t\tconst commandsToSkipShell = config.get<string[]>('commandsToSkipShell', []);\n\t\tlet updated = false;\n\t\tif (!commandsToSkipShell.includes('crowd-pilot.modelRun')) {\n\t\t\tcommandsToSkipShell.push('crowd-pilot.modelRun');\n\t\t\tupdated = true;\n\t\t}\n\t\tif (!commandsToSkipShell.includes('crowd-pilot.hideUi')) {\n\t\t\tcommandsToSkipShell.push('crowd-pilot.hideUi');\n\t\t\tupdated = true;\n\t\t}\n\t\tif (updated) {\n\t\t\tawait config.update('commandsToSkipShell', commandsToSkipShell, vscode.ConfigurationTarget.Global);\n\t\t}\n\t})().catch((err) => console.error('[crowd-pilot] Startup initialization error:', err));\n\n\tconst hideUi = vscode.commands.registerCommand('crowd-pilot.hideUi', () => {\n\t\thidePreviewUI(true);\n\t});\n\n\tconst modelRun = vscode.commands.registerCommand('crowd-pilot.modelRun', async () => {\n\t\tconst editor = vscode.window.activeTextEditor;\n\t\tif (!editor) {\n\t\t\treturn;\n\t\t}\n\t\ttry {\n\t\t\t// Confirm only when a suggestion is visible\n\t\t\tif (!previewVisible) { return; }\n\t\t\tlet action: PlannedAction | undefined = currentAction;\n\t\t\tif (!action) {\n\t\t\t\tconst single = await requestModelActions(editor);\n\t\t\t\tcurrentAction = single;\n\t\t\t\taction = single;\n\t\t\t}\n\t\t\tif (!action) {\n\t\t\t\thidePreviewUI();\n\t\t\t\treturn;\n\t\t\t}\n\t\t\thidePreviewUI(false);\n\t\t\tawait executeAction(action);\n\t\t\tautoShowNextAction();\n\t\t} catch (err) {\n\t\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\tvscode.window.showErrorMessage(`Model run failed: ${errorMessage}`);\n\t\t}\n\t});\n\n\tconst sglangTest = vscode.commands.registerCommand('crowd-pilot.sglangTest', async () => {\n\t\ttry {\n\t\t\tawait callSGLangChat();\n\t\t} catch (err) {\n\t\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\tvscode.window.showErrorMessage(`SGLang test failed: ${errorMessage}`);\n\t\t}\n\t});\n\n\t// Auto-preview listeners\n\tconst onSelChange = vscode.window.onDidChangeTextEditorSelection((e) => {\n\t\tif (e.textEditor === vscode.window.activeTextEditor) {\n\t\t\tsuppressAutoPreview = false;\n\t\t\tautoShowNextAction();\n\t\t}\n\t});\n\tconst onActiveChange = vscode.window.onDidChangeActiveTextEditor(() => {\n\t\tsuppressAutoPreview = false;\n\t\tautoShowNextAction();\n\t});\n\tconst onDocChange = vscode.workspace.onDidChangeTextDocument((e) => {\n\t\tif (vscode.window.activeTextEditor?.document === e.document) {\n\t\t\tsuppressAutoPreview = false;\n\t\t\tautoShowNextAction();\n\t\t}\n\t});\n\n\tcontext.subscriptions.push(hideUi, sglangTest, modelRun, onSelChange, onActiveChange, onDocChange);\n}\n\nexport function deactivate() {}\n\n// -------------------- Plan Types & Execution --------------------\ntype PlannedAction =\n| { kind: 'showTextDocument' }\n| { kind: 'setSelections', selections: Array<{ start: [number, number], end: [number, number] }> }\n| { kind: 'editInsert', position: [number, number], text: string }\n| { kind: 'editDelete', range: { start: [number, number], end: [number, number] } }\n| { kind: 'editReplace', range: { start: [number, number], end: [number, number] }, text: string }\n| { kind: 'terminalShow' }\n| { kind: 'terminalSendText', text: string };\n\nlet currentAction: PlannedAction | undefined;\n\nasync function executeAction(action: PlannedAction): Promise<void> {\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\tconst doc = editor.document;\n\tconst term = vscode.window.terminals[0] ?? vscode.window.createTerminal('Test');\n\tif (action.kind === 'showTextDocument') {\n\t\tawait vscode.window.showTextDocument(doc);\n\t\treturn;\n\t}\n\tif (action.kind === 'setSelections') {\n\t\teditor.selections = action.selections.map(s => new vscode.Selection(\n\t\t\tnew vscode.Position(s.start[0], s.start[1]),\n\t\t\tnew vscode.Position(s.end[0], s.end[1])\n\t\t));\n\t\tif (editor.selections.length > 0) {\n\t\t\teditor.revealRange(editor.selections[0], vscode.TextEditorRevealType.InCenterIfOutsideViewport);\n\t\t}\n\t\treturn;\n\t}\n\tif (action.kind === 'editInsert') {\n\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.insert(new vscode.Position(action.position[0], action.position[1]), action.text));\n\t\treturn;\n\t}\n\tif (action.kind === 'editDelete') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(action.range.start[0], action.range.start[1]),\n\t\t\tnew vscode.Position(action.range.end[0], action.range.end[1])\n\t\t);\n\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.delete(range));\n\t\treturn;\n\t}\n\tif (action.kind === 'editReplace') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(action.range.start[0], action.range.start[1]),\n\t\t\tnew vscode.Position(action.range.end[0], action.range.end[1])\n\t\t);\n\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.replace(range, action.text));\n\t\treturn;\n\t}\n\tif (action.kind === 'terminalShow') {\n\t\tterm.show();\n\t\treturn;\n\t}\n\tif (action.kind === 'terminalSendText') {\n\t\tterm.sendText(action.text);\n\t\treturn;\n\t}\n}\n\n// -------------------- UI State & Helpers --------------------\nconst UI_CONTEXT_KEY = 'crowdPilot.uiVisible';\nlet previewVisible = false;\nlet decorationDeleteType: vscode.TextEditorDecorationType | undefined;\nlet decorationReplaceType: vscode.TextEditorDecorationType | undefined;\nlet decorationReplaceBlockType: vscode.TextEditorDecorationType | undefined;\nlet mockStep = 0;\nlet suppressAutoPreview = false;\n\nfunction disposePreviewDecorations() {\n\ttry { decorationDeleteType?.dispose(); } catch {}\n\ttry { decorationReplaceType?.dispose(); } catch {}\n\ttry { decorationReplaceBlockType?.dispose(); } catch {}\n\tdecorationDeleteType = undefined;\n\tdecorationReplaceType = undefined;\n\tdecorationReplaceBlockType = undefined;\n}\n\nfunction getDynamicMargin(editor: vscode.TextEditor, anchorLine: number, text: string): string {\n\t// Count lines in the preview text\n\tconst lines = text.split(/\r?\n/);\n\tconst height = lines.length;\n\t\n\t// We need to check the document lines that will be covered by this panel.\n\t// The panel starts at 'anchorLine' and extends downwards by 'height' lines.\n\t// However, visually, since it's 'after', it sits to the right of 'anchorLine',\n\t// and then flows down.\n\t// So we check document lines from anchorLine to anchorLine + height - 1.\n\t\n\tconst doc = editor.document;\n\tlet maxLen = 0;\n\tconst startLine = anchorLine;\n\tconst endLine = Math.min(doc.lineCount - 1, anchorLine + height - 1);\n\t\n\tfor (let i = startLine; i <= endLine; i++) {\n\t\tconst lineText = doc.lineAt(i).text;\n\t\t// Simple approximation: assume tabs are 4 spaces if we can't get config easily, \n\t\t// or just treat them as 1 char (which might underestimate). \n\t\t// Better to overestimate: treat tab as 4 chars.\n\t\tconst len = lineText.replace(/\t/g, '    ').length;\n\t\tif (len > maxLen) {\n\t\t\tmaxLen = len;\n\t\t}\n\t}\n\t\n\t// Length of the anchor line itself\n\tconst anchorLineText = doc.lineAt(anchorLine).text;\n\tconst anchorLen = anchorLineText.replace(/\t/g, '    ').length;\n\t\n\t// The offset needed is maxLen - anchorLen.\n\t// If maxLen <= anchorLen, offset is 0 (margin is just base padding).\n\t// If maxLen > anchorLen, we need to push right by (maxLen - anchorLen).\n\t\n\tconst diff = Math.max(0, maxLen - anchorLen);\n\t// Base margin 2rem is roughly 4ch. Let's use ch units for everything to be consistent.\n\t// 1ch is width of '0'. In monospace, mostly consistent.\n\t// Add 3ch extra padding for safety/visual gap.\n\tconst margin = diff + 4; \n\treturn `${margin}ch`;\n}\n\nfunction showPreviewUI(action: PlannedAction): void {\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\tdisposePreviewDecorations();\n\n\t// FIXME (f.srambical): add file switch \n\tconst next = (action.kind === 'editInsert' || action.kind === 'editDelete' || action.kind === 'editReplace' || action.kind === 'terminalSendText' || action.kind === 'setSelections') ? action : undefined;\n\tif (!next) {\n\t\tpreviewVisible = false;\n\t\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n\t\tcurrentAction = action;\n\t\treturn;\n\t}\n\n\tconst trimText = (t: string) => {\n\t\tconst oneLine = t.replace(/\r?\n/g, '\\n');\n\t\treturn oneLine.length > 80 ? oneLine.slice(0, 77) + '…' : oneLine;\n\t};\n\n\tif (next.kind === 'setSelections') {\n\t\t// For setSelections, we only preview the primary selection's start/active position\n\t\tconst selection = next.selections[0];\n\t\tconst targetPos = new vscode.Position(selection.start[0], selection.start[1]);\n\t\t// Check if the target position is visible\n\t\tconst isVisible = editor.visibleRanges.some(r => r.contains(targetPos));\n\t\t\n\t\tlet anchorPos = targetPos;\n\t\tlet label = ""↳ Move Cursor Here"";\n\n\t\tif (!isVisible && editor.visibleRanges.length > 0) {\n\t\t\tconst firstVisible = editor.visibleRanges[0].start;\n\t\t\tconst lastVisible = editor.visibleRanges[editor.visibleRanges.length - 1].end;\n\t\t\t\n\t\t\tif (targetPos.isBefore(firstVisible)) {\n\t\t\t\tanchorPos = editor.document.lineAt(firstVisible.line).range.end;\n\t\t\t} else {\n\t\t\t\tanchorPos = editor.document.lineAt(lastVisible.line).range.end;\n\t\t\t}\n\n\t\t\tif (targetPos.line < anchorPos.line) {\n\t\t\t\tlabel = `↑ Move Cursor to Line ${targetPos.line + 1}`;\n\t\t\t} else {\n\t\t\t\tlabel = `↓ Move Cursor to Line ${targetPos.line + 1}`;\n\t\t\t}\n\t\t}\n\n\t\tconst margin = getDynamicMargin(editor, anchorPos.line, label);\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tafter: {\n\t\t\t\tcontentText: '',\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""${label}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n\t} else if (next.kind === 'terminalSendText') {\n\t\tconst cursor = editor.selection.active;\n\t\tconst lineEnd = editor.document.lineAt(cursor.line).range.end;\n\t\tconst cmd = next.text.replace(/""/g, '\\""').replace(/\r?\n/g, '\\A ');\n\t\tconst margin = getDynamicMargin(editor, cursor.line, ""↳ Execute in Terminal:\n"" + next.text);\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tafter: {\n\t\t\t\tcontentText: '',\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""↳ Execute in Terminal:\\A ${cmd}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(lineEnd, lineEnd) }]);\n\t} else if (next.kind === 'editInsert') {\n\t\tconst posLine = next.position[0];\n\t\tconst fullBlock = next.text;\n\t\tconst cssContent = fullBlock\n\t\t\t.replace(/""/g, '\\""')\n\t\t\t.replace(/\r?\n/g, '\\A ');\n\n\t\tconst docLineCount = editor.document.lineCount;\n\t\t// If inserting at EOF (or beyond), attach to the last line.\n\t\t// Otherwise, attach to the line AT the insertion point and shift visually UP into the gap.\n\t\tlet anchorLine = posLine;\n\t\tlet shiftUp = true;\n\t\t\n\t\tif (anchorLine >= docLineCount) {\n\t\t\tanchorLine = docLineCount - 1;\n\t\t\tshiftUp = false; // At EOF, we just append below or to the right\n\t\t}\n\n\t\tconst anchorPos = new vscode.Position(anchorLine, Number.MAX_VALUE); \n\t\t\n\t\t// We attach to the line AT the insertion point.\n\t\t// The panel floats to the right of this line.\n\t\t// The dashed line connects the start of this line to the panel.\n\t\t// This indicates that the new text will be inserted at this line position (pushing the current line down).\n\t\tconst marginCheckLine = anchorLine;\n\t\tconst margin = getDynamicMargin(editor, marginCheckLine, fullBlock);\n\n\t\tconst topOffset = '0';\n\n\t\t// Dashed line style\n\t\t// We use 'before' decoration for the line.\n\t\t// It needs to be absolute, full width (or enough to reach left), \n\t\t// and aligned with the panel top.\n\t\tconst beforeDecoration = {\n\t\t\tcontentText: '',\n\t\t\ttextDecoration: `none; position: absolute; left: 0; width: 100vw; border-top: 1px dashed var(--vscode-charts-purple); top: 0; height: 0; z-index: 99; pointer-events: none;`\n\t\t};\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tbefore: beforeDecoration,\n\t\t\tafter: {\n\t\t\t\tcontentText: '',\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""${cssContent}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top; top: ${topOffset};`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n\t} else if (next.kind === 'editDelete') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(next.range.start[0], next.range.start[1]),\n\t\t\tnew vscode.Position(next.range.end[0], next.range.end[1])\n\t\t);\n\t\tdecorationDeleteType = vscode.window.createTextEditorDecorationType({\n\t\t\tbackgroundColor: 'rgba(255, 60, 60, 0.18)',\n\t\t\tborder: '1px solid rgba(255, 60, 60, 0.35)',\n\t\t\ttextDecoration: 'line-through'\n\t\t});\n\t\teditor.setDecorations(decorationDeleteType, [{ range }]);\n\t} else if (next.kind === 'editReplace') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(next.range.start[0], next.range.start[1]),\n\t\t\tnew vscode.Position(next.range.end[0], next.range.end[1])\n\t\t);\n\t\t// Highlight original range (to be replaced)\n\t\tdecorationReplaceType = vscode.window.createTextEditorDecorationType({\n\t\t\tbackgroundColor: 'rgba(255,165,0,0.15)',\n\t\t\tborder: '1px dashed rgba(255,165,0,0.45)',\n\t\t\tcolor: new vscode.ThemeColor('disabledForeground'),\n\t\t\ttextDecoration: 'line-through'\n\t\t});\n\t\teditor.setDecorations(decorationReplaceType, [{ range }]);\n\n\t\t// Show replacement block to the right of the first replaced line\n\t\tconst fullBlock = next.text;\n\t\t\n\t\t// CSS-escape the text for the 'content' property:\n\t\t// - Escape double quotes\n\t\t// - Replace newlines with \A (CSS newline)\n\t\tconst cssContent = fullBlock\n\t\t\t.replace(/""/g, '\\""')\n\t\t\t.replace(/\r?\n/g, '\\A '); \n\n\t\t// Attach 'after' decoration to the start of the replacement range\n\t\t// (Actually, attaching to the end of the first line is safer for 'after')\n\t\tconst anchorLine = range.start.line;\n\t\tconst anchorPos = new vscode.Position(anchorLine, Number.MAX_VALUE);\n\t\tconst margin = getDynamicMargin(editor, anchorLine, fullBlock);\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tafter: {\n\t\t\t\tcontentText: '', // Handled by CSS content\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""${cssContent}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n\t}\n\n\tpreviewVisible = true;\n\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, true);\n\tcurrentAction = action;\n}\n\nfunction hidePreviewUI(suppress?: boolean): void {\n\tdisposePreviewDecorations();\n\tpreviewVisible = false;\n\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n\tif (suppress) {\n\t\tsuppressAutoPreview = true;\n\t}\n}\n\n// -------------------- Hardcoded single-step actions --------------------\nfunction getHardcodedNextAction(editor: vscode.TextEditor): PlannedAction | undefined {\n\tconst cursor = editor.selection.active;\n\tconst doc = editor.document;\n\tconst lineCount = doc.lineCount;\n\tconst clamp = (n: number, min: number, max: number) => Math.max(min, Math.min(max, n));\n\n\t// Step 0: Insert multiline content two lines below the cursor (start of target line)\n\tif (mockStep === 0) {\n\t\tconst targetLine = clamp(cursor.line + 2, 0, Math.max(0, lineCount - 1));\n\t\treturn {\n\t\t\tkind: 'editInsert',\n\t\t\tposition: [targetLine, 0],\n\t\t\ttext: '/* crowd-pilot: insert start */\nline A\nline B\n/* crowd-pilot: insert end */\n'\n\t\t};\n\t}\n\t// Step 1: Replace a two-line range three and four lines below the cursor\n\tif (mockStep === 1) {\n\t\tconst startLine = clamp(cursor.line + 3, 0, Math.max(0, lineCount - 1));\n\t\tconst endLine = clamp(startLine + 1, 0, Math.max(0, lineCount - 1));\n\t\tconst endChar = doc.lineAt(endLine).range.end.character;\n\t\tconst range = {\n\t\t\tstart: [startLine, 0] as [number, number],\n\t\t\tend: [endLine, endChar] as [number, number]\n\t\t};\n\t\tconst replacement = [\n\t\t\t'/* crowd-pilot: replacement */',\n\t\t\t'REPLACED LINE 1',\n\t\t\t'REPLACED LINE 2'\n\t\t].join('\n');\n\t\treturn { kind: 'editReplace', range, text: replacement };\n\t}\n\t// Step 2: Delete a three-line range six to eight lines below the cursor\n\tif (mockStep === 2) {\n\t\tconst startLine = clamp(cursor.line + 6, 0, Math.max(0, lineCount - 1));\n\t\tconst endLine = clamp(startLine + 2, 0, Math.max(0, lineCount - 1));\n\t\t\n\t\t// To fully delete the lines including the newline, we target the start of the next line.\n\t\tlet endPosLine = endLine + 1;\n\t\tlet endPosChar = 0;\n\t\t\n\t\tif (endPosLine >= lineCount) {\n\t\t\t// If deleting the last line(s), just go to the end of the document\n\t\t\tendPosLine = lineCount - 1;\n\t\t\tendPosChar = doc.lineAt(endPosLine).range.end.character;\n\t\t}\n\n\t\tconst range = {\n\t\t\tstart: [startLine, 0] as [number, number],\n\t\t\tend: [endPosLine, endPosChar] as [number, number]\n\t\t};\n\t\treturn { kind: 'editDelete', range };\n\t}\n\t// Step 3: Execute in Terminal\n\tif (mockStep === 3) {\n\t\treturn { kind: 'terminalSendText', text: 'echo ""Hello World""' };\n\t}\n\t// Step 4: Move Cursor to End of File\n\tif (mockStep === 4) {\n\t\tconst lastLine = doc.lineCount - 1;\n\t\tconst lastChar = doc.lineAt(lastLine).range.end.character;\n\t\treturn {\n\t\t\tkind: 'setSelections',\n\t\t\tselections: [{ start: [lastLine, lastChar], end: [lastLine, lastChar] }]\n\t\t};\n\t}\n\treturn undefined;\n}\n\nfunction advanceMockStep(): void {\n\tmockStep = (mockStep + 1) % 5;\n}\n\nasync function autoShowNextAction(): Promise<void> {\n\tif (suppressAutoPreview) { return; }\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\ttry {\n\t\tconst next = await requestModelActions(editor);\n\t\tif (next) {\n\t\t\tshowPreviewUI(next);\n\t\t} else {\n\t\t\thidePreviewUI();\n\t\t}\n\t} catch (err) {\n\t\thidePreviewUI();\n\t}\n}\n\n// -------------------- SGLang Client (simple test) --------------------\nasync function callSGLangChat(): Promise<void> {\n\tconst requestBody = {\n\t\tmodel: 'qwen/qwen2.5-0.5b-instruct',\n\t\tmessages: [\n\t\t\t{ role: 'user', content: 'What is the capital of France?' }\n\t\t]\n\t};\n\tconst postData = JSON.stringify(requestBody);\n\n\tconst options = {\n\t\thostname: HOSTNAME,\n\t\tport: PORT,\n\t\tpath: '/v1/chat/completions',\n\t\tmethod: 'POST',\n\t\theaders: {\n\t\t\t'Content-Type': 'application/json',\n\t\t\t'Content-Length': Buffer.byteLength(postData)\n\t\t}\n\t};\n\n\ttry {\n\t\tconst json = await new Promise<any>((resolve, reject) => {\n\t\t\tconst req = http.request(options, (res: http.IncomingMessage) => {\n\t\t\t\tlet data = '';\n\t\t\t\tres.on('data', (chunk: Buffer) => {\n\t\t\t\t\tdata += chunk.toString();\n\t\t\t\t});\n\t\t\t\tres.on('end', () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tresolve(JSON.parse(data));\n\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\treject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t});\n\n\t\t\treq.on('error', (err: Error) => {\n\t\t\t\treject(err);\n\t\t\t});\n\n\t\t\treq.write(postData);\n\t\t\treq.end();\n\t\t});\n\n\t\tvscode.window.showInformationMessage(`SGLang response: ${JSON.stringify(json, null, 2)}`);\n\t} catch (err) {\n\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\tvscode.window.showErrorMessage(`SGLang request failed: ${errorMessage}`);\n\t}\n}\n\n// -------------------- Model-planned Actions --------------------\nasync function requestModelActions(editor: vscode.TextEditor): Promise<PlannedAction> {\n\tconst schemaDescription = [\n\t\t'Role: You suggest the next VS Code editor/terminal action to progress the current task.',\n\t\t'Output ONLY a JSON object (no prose, no code fences).',\n\t\t'Coordinates are zero-based [line, column].',\n\t\t'Allowed actions (JSON schema-like):',\n\t\t'{ kind: ""sowTextDocument"" }',\n\t\t'{ kind: ""setSelections"", selections: Array<{ start: [number, number], end: [number, number] }> }',\n\t\t'{ kind: ""editInsert"", position: [number, number], text: string }',\n\t\t'{ kind: ""editDelete"", range: { start: [number, number], end: [number, number] } }',\n\t\t'{ kind: ""editReplace"", range: { start: [number, number], end: [number, number] }, text: string }',\n\t\t'{ kind: ""terminalShow"" }',\n\t\t'{ kind: ""terminalSendText"", text: string }',\n\t\t'Guidelines:',\n\t\t'- If you you insert text, insert until the logical end of the current statement or block.',\n\t\t'- When inserting text, make sure to not repeat existing text (except when replacing existing text).',\n\t\t'- Use double-quoted JSON strings.'\n\t].join('\n');\n\n\tconst doc = editor.document;\n\tconst cursor = editor.selection.active;\n\tconst fullText = doc.getText();\n\tconst numberedContext = fullText.split(/\r?\n/).map((line, i) => `${i}: ${line}`).join('\n');\n\n\tconst tabbingPrompt = [\n\t\t'Your role: Propose the single next action according to the schema to help the developer progress.',\n\t\t'',\n\t\t'Available context:',\n\t\t`- File: ${doc.fileName}`,\n\t\t`- Language: ${doc.languageId}`,\n\t\t`- Cursor: (${cursor.line}, ${cursor.character})`,\n\t\t'',\n\t\t'Full file content (zero-based line numbers):',\n\t\t'```',\n\t\tnumberedContext,\n\t\t'```',\n\t\t'',\n\t\t'Respond with ONLY a JSON object containing exactly one action.'\n\t].join('\n');\n\n\tconst requestBody = {\n\t\tmodel: 'qwen/qwen2.5-0.5b-instruct',\n\t\tmessages: [\n\t\t\t{ role: 'system', content: schemaDescription },\n\t\t\t{ role: 'user', content: tabbingPrompt }\n\t\t]\n\t};\n\n\tconst postData = JSON.stringify(requestBody);\n\tconst options = {\n\t\thostname: HOSTNAME,\n\t\tport: PORT,\n\t\tpath: '/v1/chat/completions',\n\t\tmethod: 'POST',\n\t\theaders: {\n\t\t\t'Content-Type': 'application/json',\n\t\t\t'Content-Length': Buffer.byteLength(postData)\n\t\t}\n\t};\n\n\tconst json = await new Promise<any>((resolve, reject) => {\n\t\tconst req = http.request(options, (res: http.IncomingMessage) => {\n\t\t\tlet data = '';\n\t\t\tres.on('data', (chunk: Buffer) => { data += chunk.toString(); });\n\t\t\tres.on('end', () => {\n\t\t\t\ttry {\n\t\t\t\t\tresolve(JSON.parse(data));\n\t\t\t\t} catch (err) {\n\t\t\t\t\treject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n\t\t\t\t}\n\t\t\t});\n\t\t});\n\t\treq.on('error', (err: Error) => reject(err));\n\t\treq.write(postData);\n\t\treq.end();\n\t});\n\n\tconst content = extractChatContent(json);\n\tif (typeof content !== 'string' || content.trim().length === 0) {\n\t\tthrow new Error('Empty model content');\n\t}\n\tconst action = parsePlannedAction(content);\n\tif (!action) {\n\t\tthrow new Error('No valid action parsed from model output');\n\t}\n\treturn action;\n}\n\nfunction extractChatContent(json: any): string | undefined {\n\ttry {\n\t\tif (json && Array.isArray(json.choices) && json.choices[0]) {\n\t\t\tconst choice = json.choices[0];\n\t\t\tif (choice.message && typeof choice.message.content === 'string') {\n\t\t\t\treturn choice.message.content;\n\t\t\t}\n\t\t\tif (typeof choice.text === 'string') {\n\t\t\t\treturn choice.text;\n\t\t\t}\n\t\t}\n\t\treturn undefined;\n\t} catch {\n\t\treturn undefined;\n\t}\n}\n\nfunction parsePlannedAction(raw: string): PlannedAction | undefined {\n\tlet text = raw.trim();\n\ttext = text.replace(/^```(?:json)?\s*/i, '').replace(/```\s*$/i, '').trim();\n\ttext = text.replace(/<think>[\s\S]*?<\/think>/gi, '').trim();\n\tlet parsed: any;\n\ttry {\n\t\tparsed = JSON.parse(text);\n\t} catch (err) {\n\t\treturn undefined;\n\t}\n\tif (Array.isArray(parsed)) {\n\t\tconsole.error('Model should not return an array.');\n\t\treturn undefined;\n\t}\n\tswitch (parsed.kind) {\n\t\tcase 'showTextDocument':\n\t\t\treturn { kind: 'showTextDocument' };\n\t\tcase 'setSelections': {\n\t\t\tconst selections = Array.isArray(parsed.selections) ? parsed.selections : [];\n\t\t\tconst norm = selections.map((s: any) => ({\n\t\t\t\tstart: Array.isArray(s?.start) && s.start.length === 2 ? [Number(s.start[0]) || 0, Number(s.start[1]) || 0] as [number, number] : [0, 0] as [number, number],\n\t\t\t\tend: Array.isArray(s?.end) && s.end.length === 2 ? [Number(s.end[0]) || 0, Number(s.end[1]) || 0] as [number, number] : [0, 0] as [number, number]\n\t\t\t}));\n\t\t\treturn { kind: 'setSelections', selections: norm };\n\t\t}\n\t\tcase 'editInsert': {\n\t\t\tconst pos = Array.isArray(parsed.position) && parsed.position.length === 2 ? [Number(parsed.position[0]) || 0, Number(parsed.position[1]) || 0] as [number, number] : [0, 0] as [number, number];\n\t\t\tconst textVal = typeof parsed.text === 'string' ? parsed.text : '';\n\t\t\treturn { kind: 'editInsert', position: pos, text: textVal };\n\t\t}\n\t\tcase 'editDelete': {\n\t\t\tconst start = Array.isArray(parsed.range?.start) && parsed.range.start.length === 2 ? [Number(parsed.range.start[0]) || 0, Number(parsed.range.start[1]) || 0] as [number, number] : [0, 0] as [number, number];\n\t\t\tconst end = Array.isArray(parsed.range?.end) && parsed.range.end.length === 2 ? [Number(parsed.range.end[0]) || 0, Number(parsed.range.end[1]) || 0] as [number, number] : [0, 0] as [number, number];\n\t\t\treturn { kind: 'editDelete', range: { start, end } };\n\t\t}\n\t\tcase 'editReplace': {\n\t\t\tconst start = Array.isArray(parsed.range?.start) && parsed.range.start.length === 2 ? [Number(parsed.range.start[0]) || 0, Number(parsed.range.start[1]) || 0] as [number, number] : [0, 0] as [number, number];\n\t\t\tconst end = Array.isArray(parsed.range?.end) && parsed.range.end.length === 2 ? [Number(parsed.range.end[0]) || 0, Number(parsed.range.end[1]) || 0] as [number, number] : [0, 0] as [number, number];\n\t\t\tconst textVal = typeof parsed.text === 'string' ? parsed.text : '';\n\t\t\treturn { kind: 'editReplace', range: { start, end }, text: textVal };\n\t\t}\n\t\tcase 'terminalShow':\n\t\t\treturn { kind: 'terminalShow' };\n\t\tcase 'terminalSendText': {\n\t\t\tconst textVal = typeof parsed.text === 'string' ? parsed.text : '';\n\t\t\treturn { kind: 'terminalSendText', text: textVal };\n\t\t}\n\t\tdefault:\n\t\t\treturn undefined;\n\t}\n}",typescript,tab
+2,299,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"3:35:04 PM [info] Activating crowd-code\n3:35:04 PM [info] Recording started\n3:35:04 PM [info] Initializing git provider using file system watchers...\n3:35:05 PM [info] Git repository found\n3:35:05 PM [info] Git provider initialized successfully\n3:35:05 PM [info] Initial git state: [object Object]\n",Log,tab
+3,5984208,"src/extension.ts",0,0,"",typescript,tab
+4,5992260,"src/extension.ts",34,22785,"import * as https from 'https';\nimport * as http from 'http';\nimport { Buffer } from 'buffer';\n\n\nconst SGLANG_HOSTNAME = 'hai001';\nconst SGLANG_PORT = 30000;\nconst SGLANG_BASE_PATH = '/v1/chat/completions';\nconst SGLANG_MODEL_NAME = 'qwen/qwen3-0.6b';\n\nconst GEMINI_HOSTNAME = 'generativelanguage.googleapis.com';\nconst GEMINI_PORT = 443;\nconst GEMINI_BASE_PATH = '/v1beta/openai/chat/completions';\nconst GEMINI_MODEL_NAME = 'gemini-2.5-flash';\n\nconst USE_GEMINI = true;\n\nexport function activate(context: vscode.ExtensionContext) {\n\n\tconsole.log('[crowd-pilot] Extension activated');\n\n\t// Configure terminal to allow tab keybinding to work\n\t(async () => {\n\t\tconst config = vscode.workspace.getConfiguration('terminal.integrated');\n\t\tconst commandsToSkipShell = config.get<string[]>('commandsToSkipShell', []);\n\t\tlet updated = false;\n\t\tif (!commandsToSkipShell.includes('crowd-pilot.modelRun')) {\n\t\t\tcommandsToSkipShell.push('crowd-pilot.modelRun');\n\t\t\tupdated = true;\n\t\t}\n\t\tif (!commandsToSkipShell.includes('crowd-pilot.hideUi')) {\n\t\t\tcommandsToSkipShell.push('crowd-pilot.hideUi');\n\t\t\tupdated = true;\n\t\t}\n\t\tif (updated) {\n\t\t\tawait config.update('commandsToSkipShell', commandsToSkipShell, vscode.ConfigurationTarget.Global);\n\t\t}\n\t})().catch((err) => console.error('[crowd-pilot] Startup initialization error:', err));\n\n\tconst hideUi = vscode.commands.registerCommand('crowd-pilot.hideUi', () => {\n\t\thidePreviewUI(true);\n\t});\n\n\tconst modelRun = vscode.commands.registerCommand('crowd-pilot.modelRun', async () => {\n\t\tconst editor = vscode.window.activeTextEditor;\n\t\tif (!editor) {\n\t\t\treturn;\n\t\t}\n\t\ttry {\n\t\t\t// Confirm only when a suggestion is visible\n\t\t\tif (!previewVisible) { return; }\n\t\t\tlet action: PlannedAction | undefined = currentAction;\n\t\t\tif (!action) {\n\t\t\t\tconst single = await requestModelActions(editor);\n\t\t\t\tcurrentAction = single;\n\t\t\t\taction = single;\n\t\t\t}\n\t\t\tif (!action) {\n\t\t\t\thidePreviewUI();\n\t\t\t\treturn;\n\t\t\t}\n\t\t\thidePreviewUI(false);\n\t\t\tawait executeAction(action);\n\t\t\tautoShowNextAction();\n\t\t} catch (err) {\n\t\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\tvscode.window.showErrorMessage(`Model run failed: ${errorMessage}`);\n\t\t}\n\t});\n\n\tconst sglangTest = vscode.commands.registerCommand('crowd-pilot.sglangTest', async () => {\n\t\ttry {\n\t\t\tawait callSGLangChat();\n\t\t} catch (err) {\n\t\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\t\tvscode.window.showErrorMessage(`SGLang test failed: ${errorMessage}`);\n\t\t}\n\t});\n\n\t// Auto-preview listeners\n\tconst onSelChange = vscode.window.onDidChangeTextEditorSelection((e) => {\n\t\tif (e.textEditor === vscode.window.activeTextEditor) {\n\t\t\tsuppressAutoPreview = false;\n\t\t\tautoShowNextAction();\n\t\t}\n\t});\n\tconst onActiveChange = vscode.window.onDidChangeActiveTextEditor(() => {\n\t\tsuppressAutoPreview = false;\n\t\tautoShowNextAction();\n\t});\n\tconst onDocChange = vscode.workspace.onDidChangeTextDocument((e) => {\n\t\tif (vscode.window.activeTextEditor?.document === e.document) {\n\t\t\tsuppressAutoPreview = false;\n\t\t\tautoShowNextAction();\n\t\t}\n\t});\n\n\tcontext.subscriptions.push(hideUi, sglangTest, modelRun, onSelChange, onActiveChange, onDocChange);\n}\n\nexport function deactivate() {}\n\n// -------------------- Plan Types & Execution --------------------\ntype PlannedAction =\n| { kind: 'showTextDocument' }\n| { kind: 'setSelections', selections: Array<{ start: [number, number], end: [number, number] }> }\n| { kind: 'editInsert', position: [number, number], text: string }\n| { kind: 'editDelete', range: { start: [number, number], end: [number, number] } }\n| { kind: 'editReplace', range: { start: [number, number], end: [number, number] }, text: string }\n| { kind: 'terminalShow' }\n| { kind: 'terminalSendText', text: string };\n\nlet currentAction: PlannedAction | undefined;\n\nasync function executeAction(action: PlannedAction): Promise<void> {\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\tconst doc = editor.document;\n\tconst term = vscode.window.terminals[0] ?? vscode.window.createTerminal('Test');\n\tif (action.kind === 'showTextDocument') {\n\t\tawait vscode.window.showTextDocument(doc);\n\t\treturn;\n\t}\n\tif (action.kind === 'setSelections') {\n\t\teditor.selections = action.selections.map(s => new vscode.Selection(\n\t\t\tnew vscode.Position(s.start[0], s.start[1]),\n\t\t\tnew vscode.Position(s.end[0], s.end[1])\n\t\t));\n\t\tif (editor.selections.length > 0) {\n\t\t\teditor.revealRange(editor.selections[0], vscode.TextEditorRevealType.InCenterIfOutsideViewport);\n\t\t}\n\t\treturn;\n\t}\n\tif (action.kind === 'editInsert') {\n\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.insert(new vscode.Position(action.position[0], action.position[1]), action.text));\n\t\treturn;\n\t}\n\tif (action.kind === 'editDelete') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(action.range.start[0], action.range.start[1]),\n\t\t\tnew vscode.Position(action.range.end[0], action.range.end[1])\n\t\t);\n\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.delete(range));\n\t\treturn;\n\t}\n\tif (action.kind === 'editReplace') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(action.range.start[0], action.range.start[1]),\n\t\t\tnew vscode.Position(action.range.end[0], action.range.end[1])\n\t\t);\n\t\tawait editor.edit((e: vscode.TextEditorEdit) => e.replace(range, action.text));\n\t\treturn;\n\t}\n\tif (action.kind === 'terminalShow') {\n\t\tterm.show();\n\t\treturn;\n\t}\n\tif (action.kind === 'terminalSendText') {\n\t\tterm.sendText(action.text);\n\t\treturn;\n\t}\n}\n\n// -------------------- UI State & Helpers --------------------\nconst UI_CONTEXT_KEY = 'crowdPilot.uiVisible';\nlet previewVisible = false;\nlet decorationDeleteType: vscode.TextEditorDecorationType | undefined;\nlet decorationReplaceType: vscode.TextEditorDecorationType | undefined;\nlet decorationReplaceBlockType: vscode.TextEditorDecorationType | undefined;\nlet mockStep = 0;\nlet suppressAutoPreview = false;\n\nfunction disposePreviewDecorations() {\n\ttry { decorationDeleteType?.dispose(); } catch {}\n\ttry { decorationReplaceType?.dispose(); } catch {}\n\ttry { decorationReplaceBlockType?.dispose(); } catch {}\n\tdecorationDeleteType = undefined;\n\tdecorationReplaceType = undefined;\n\tdecorationReplaceBlockType = undefined;\n}\n\nfunction getDynamicMargin(editor: vscode.TextEditor, anchorLine: number, text: string): string {\n\t// Count lines in the preview text\n\tconst lines = text.split(/\r?\n/);\n\tconst height = lines.length;\n\t\n\t// We need to check the document lines that will be covered by this panel.\n\t// The panel starts at 'anchorLine' and extends downwards by 'height' lines.\n\t// However, visually, since it's 'after', it sits to the right of 'anchorLine',\n\t// and then flows down.\n\t// So we check document lines from anchorLine to anchorLine + height - 1.\n\t\n\tconst doc = editor.document;\n\tlet maxLen = 0;\n\tconst startLine = anchorLine;\n\tconst endLine = Math.min(doc.lineCount - 1, anchorLine + height - 1);\n\t\n\tfor (let i = startLine; i <= endLine; i++) {\n\t\tconst lineText = doc.lineAt(i).text;\n\t\t// Simple approximation: assume tabs are 4 spaces if we can't get config easily, \n\t\t// or just treat them as 1 char (which might underestimate). \n\t\t// Better to overestimate: treat tab as 4 chars.\n\t\tconst len = lineText.replace(/\t/g, '    ').length;\n\t\tif (len > maxLen) {\n\t\t\tmaxLen = len;\n\t\t}\n\t}\n\t\n\t// Length of the anchor line itself\n\tconst anchorLineText = doc.lineAt(anchorLine).text;\n\tconst anchorLen = anchorLineText.replace(/\t/g, '    ').length;\n\t\n\t// The offset needed is maxLen - anchorLen.\n\t// If maxLen <= anchorLen, offset is 0 (margin is just base padding).\n\t// If maxLen > anchorLen, we need to push right by (maxLen - anchorLen).\n\t\n\tconst diff = Math.max(0, maxLen - anchorLen);\n\t// Base margin 2rem is roughly 4ch. Let's use ch units for everything to be consistent.\n\t// 1ch is width of '0'. In monospace, mostly consistent.\n\t// Add 3ch extra padding for safety/visual gap.\n\tconst margin = diff + 4; \n\treturn `${margin}ch`;\n}\n\nfunction showPreviewUI(action: PlannedAction): void {\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\tdisposePreviewDecorations();\n\n\t// FIXME (f.srambical): add file switch \n\tconst next = (action.kind === 'editInsert' || action.kind === 'editDelete' || action.kind === 'editReplace' || action.kind === 'terminalSendText' || action.kind === 'setSelections') ? action : undefined;\n\tif (!next) {\n\t\tpreviewVisible = false;\n\t\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n\t\tcurrentAction = action;\n\t\treturn;\n\t}\n\n\tconst trimText = (t: string) => {\n\t\tconst oneLine = t.replace(/\r?\n/g, '\\n');\n\t\treturn oneLine.length > 80 ? oneLine.slice(0, 77) + '…' : oneLine;\n\t};\n\n\tif (next.kind === 'setSelections') {\n\t\t// For setSelections, we only preview the primary selection's start/active position\n\t\tconst selection = next.selections[0];\n\t\tconst targetPos = new vscode.Position(selection.start[0], selection.start[1]);\n\t\t// Check if the target position is visible\n\t\tconst isVisible = editor.visibleRanges.some(r => r.contains(targetPos));\n\t\t\n\t\tlet anchorPos = targetPos;\n\t\tlet label = ""↳ Move Cursor Here"";\n\n\t\tif (!isVisible && editor.visibleRanges.length > 0) {\n\t\t\tconst firstVisible = editor.visibleRanges[0].start;\n\t\t\tconst lastVisible = editor.visibleRanges[editor.visibleRanges.length - 1].end;\n\t\t\t\n\t\t\tif (targetPos.isBefore(firstVisible)) {\n\t\t\t\tanchorPos = editor.document.lineAt(firstVisible.line).range.end;\n\t\t\t} else {\n\t\t\t\tanchorPos = editor.document.lineAt(lastVisible.line).range.end;\n\t\t\t}\n\n\t\t\tif (targetPos.line < anchorPos.line) {\n\t\t\t\tlabel = `↑ Move Cursor to Line ${targetPos.line + 1}`;\n\t\t\t} else {\n\t\t\t\tlabel = `↓ Move Cursor to Line ${targetPos.line + 1}`;\n\t\t\t}\n\t\t}\n\n\t\tconst margin = getDynamicMargin(editor, anchorPos.line, label);\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tafter: {\n\t\t\t\tcontentText: '',\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""${label}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n\t} else if (next.kind === 'terminalSendText') {\n\t\tconst cursor = editor.selection.active;\n\t\tconst lineEnd = editor.document.lineAt(cursor.line).range.end;\n\t\tconst cmd = next.text.replace(/""/g, '\\""').replace(/\r?\n/g, '\\A ');\n\t\tconst margin = getDynamicMargin(editor, cursor.line, ""↳ Execute in Terminal:\n"" + next.text);\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tafter: {\n\t\t\t\tcontentText: '',\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""↳ Execute in Terminal:\\A ${cmd}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(lineEnd, lineEnd) }]);\n\t} else if (next.kind === 'editInsert') {\n\t\tconst posLine = next.position[0];\n\t\tconst fullBlock = next.text;\n\t\tconst cssContent = fullBlock\n\t\t\t.replace(/""/g, '\\""')\n\t\t\t.replace(/\r?\n/g, '\\A ');\n\n\t\tconst docLineCount = editor.document.lineCount;\n\t\t// If inserting at EOF (or beyond), attach to the last line.\n\t\t// Otherwise, attach to the line AT the insertion point and shift visually UP into the gap.\n\t\tlet anchorLine = posLine;\n\t\tlet shiftUp = true;\n\t\t\n\t\tif (anchorLine >= docLineCount) {\n\t\t\tanchorLine = docLineCount - 1;\n\t\t\tshiftUp = false; // At EOF, we just append below or to the right\n\t\t}\n\n\t\tconst anchorPos = new vscode.Position(anchorLine, Number.MAX_VALUE); \n\t\t\n\t\t// We attach to the line AT the insertion point.\n\t\t// The panel floats to the right of this line.\n\t\t// The dashed line connects the start of this line to the panel.\n\t\t// This indicates that the new text will be inserted at this line position (pushing the current line down).\n\t\tconst marginCheckLine = anchorLine;\n\t\tconst margin = getDynamicMargin(editor, marginCheckLine, fullBlock);\n\n\t\tconst topOffset = '0';\n\n\t\t// Dashed line style\n\t\t// We use 'before' decoration for the line.\n\t\t// It needs to be absolute, full width (or enough to reach left), \n\t\t// and aligned with the panel top.\n\t\tconst beforeDecoration = {\n\t\t\tcontentText: '',\n\t\t\ttextDecoration: `none; position: absolute; left: 0; width: 100vw; border-top: 1px dashed var(--vscode-charts-purple); top: 0; height: 0; z-index: 99; pointer-events: none;`\n\t\t};\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tbefore: beforeDecoration,\n\t\t\tafter: {\n\t\t\t\tcontentText: '',\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""${cssContent}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top; top: ${topOffset};`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n\t} else if (next.kind === 'editDelete') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(next.range.start[0], next.range.start[1]),\n\t\t\tnew vscode.Position(next.range.end[0], next.range.end[1])\n\t\t);\n\t\tdecorationDeleteType = vscode.window.createTextEditorDecorationType({\n\t\t\tbackgroundColor: 'rgba(255, 60, 60, 0.18)',\n\t\t\tborder: '1px solid rgba(255, 60, 60, 0.35)',\n\t\t\ttextDecoration: 'line-through'\n\t\t});\n\t\teditor.setDecorations(decorationDeleteType, [{ range }]);\n\t} else if (next.kind === 'editReplace') {\n\t\tconst range = new vscode.Range(\n\t\t\tnew vscode.Position(next.range.start[0], next.range.start[1]),\n\t\t\tnew vscode.Position(next.range.end[0], next.range.end[1])\n\t\t);\n\t\t// Highlight original range (to be replaced)\n\t\tdecorationReplaceType = vscode.window.createTextEditorDecorationType({\n\t\t\tbackgroundColor: 'rgba(255,165,0,0.15)',\n\t\t\tborder: '1px dashed rgba(255,165,0,0.45)',\n\t\t\tcolor: new vscode.ThemeColor('disabledForeground'),\n\t\t\ttextDecoration: 'line-through'\n\t\t});\n\t\teditor.setDecorations(decorationReplaceType, [{ range }]);\n\n\t\t// Show replacement block to the right of the first replaced line\n\t\tconst fullBlock = next.text;\n\t\t\n\t\t// CSS-escape the text for the 'content' property:\n\t\t// - Escape double quotes\n\t\t// - Replace newlines with \A (CSS newline)\n\t\tconst cssContent = fullBlock\n\t\t\t.replace(/""/g, '\\""')\n\t\t\t.replace(/\r?\n/g, '\\A '); \n\n\t\t// Attach 'after' decoration to the start of the replacement range\n\t\t// (Actually, attaching to the end of the first line is safer for 'after')\n\t\tconst anchorLine = range.start.line;\n\t\tconst anchorPos = new vscode.Position(anchorLine, Number.MAX_VALUE);\n\t\tconst margin = getDynamicMargin(editor, anchorLine, fullBlock);\n\n\t\tdecorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n\t\t\tafter: {\n\t\t\t\tcontentText: '', // Handled by CSS content\n\t\t\t\tcolor: new vscode.ThemeColor('charts.purple'),\n\t\t\t\tbackgroundColor: new vscode.ThemeColor('editor.background'),\n\t\t\t\tfontStyle: 'italic',\n\t\t\t\tfontWeight: '600',\n\t\t\t\tmargin: `0 0 0 ${margin}`,\n\t\t\t\ttextDecoration: `none; display: inline-block; white-space: pre; content: ""${cssContent}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n\t\t\t}\n\t\t});\n\t\teditor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n\t}\n\n\tpreviewVisible = true;\n\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, true);\n\tcurrentAction = action;\n}\n\nfunction hidePreviewUI(suppress?: boolean): void {\n\tdisposePreviewDecorations();\n\tpreviewVisible = false;\n\tvscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n\tif (suppress) {\n\t\tsuppressAutoPreview = true;\n\t}\n}\n\n// -------------------- Hardcoded single-step actions --------------------\nfunction getHardcodedNextAction(editor: vscode.TextEditor): PlannedAction | undefined {\n\tconst cursor = editor.selection.active;\n\tconst doc = editor.document;\n\tconst lineCount = doc.lineCount;\n\tconst clamp = (n: number, min: number, max: number) => Math.max(min, Math.min(max, n));\n\n\t// Step 0: Insert multiline content two lines below the cursor (start of target line)\n\tif (mockStep === 0) {\n\t\tconst targetLine = clamp(cursor.line + 2, 0, Math.max(0, lineCount - 1));\n\t\treturn {\n\t\t\tkind: 'editInsert',\n\t\t\tposition: [targetLine, 0],\n\t\t\ttext: '/* crowd-pilot: insert start */\nline A\nline B\n/* crowd-pilot: insert end */\n'\n\t\t};\n\t}\n\t// Step 1: Replace a two-line range three and four lines below the cursor\n\tif (mockStep === 1) {\n\t\tconst startLine = clamp(cursor.line + 3, 0, Math.max(0, lineCount - 1));\n\t\tconst endLine = clamp(startLine + 1, 0, Math.max(0, lineCount - 1));\n\t\tconst endChar = doc.lineAt(endLine).range.end.character;\n\t\tconst range = {\n\t\t\tstart: [startLine, 0] as [number, number],\n\t\t\tend: [endLine, endChar] as [number, number]\n\t\t};\n\t\tconst replacement = [\n\t\t\t'/* crowd-pilot: replacement */',\n\t\t\t'REPLACED LINE 1',\n\t\t\t'REPLACED LINE 2'\n\t\t].join('\n');\n\t\treturn { kind: 'editReplace', range, text: replacement };\n\t}\n\t// Step 2: Delete a three-line range six to eight lines below the cursor\n\tif (mockStep === 2) {\n\t\tconst startLine = clamp(cursor.line + 6, 0, Math.max(0, lineCount - 1));\n\t\tconst endLine = clamp(startLine + 2, 0, Math.max(0, lineCount - 1));\n\t\t\n\t\t// To fully delete the lines including the newline, we target the start of the next line.\n\t\tlet endPosLine = endLine + 1;\n\t\tlet endPosChar = 0;\n\t\t\n\t\tif (endPosLine >= lineCount) {\n\t\t\t// If deleting the last line(s), just go to the end of the document\n\t\t\tendPosLine = lineCount - 1;\n\t\t\tendPosChar = doc.lineAt(endPosLine).range.end.character;\n\t\t}\n\n\t\tconst range = {\n\t\t\tstart: [startLine, 0] as [number, number],\n\t\t\tend: [endPosLine, endPosChar] as [number, number]\n\t\t};\n\t\treturn { kind: 'editDelete', range };\n\t}\n\t// Step 3: Execute in Terminal\n\tif (mockStep === 3) {\n\t\treturn { kind: 'terminalSendText', text: 'echo ""Hello World""' };\n\t}\n\t// Step 4: Move Cursor to End of File\n\tif (mockStep === 4) {\n\t\tconst lastLine = doc.lineCount - 1;\n\t\tconst lastChar = doc.lineAt(lastLine).range.end.character;\n\t\treturn {\n\t\t\tkind: 'setSelections',\n\t\t\tselections: [{ start: [lastLine, lastChar], end: [lastLine, lastChar] }]\n\t\t};\n\t}\n\treturn undefined;\n}\n\nfunction advanceMockStep(): void {\n\tmockStep = (mockStep + 1) % 5;\n}\n\nasync function autoShowNextAction(): Promise<void> {\n\tif (suppressAutoPreview) { return; }\n\tconst editor = vscode.window.activeTextEditor;\n\tif (!editor) { return; }\n\ttry {\n\t\tconst next = await requestModelActions(editor);\n\t\tif (next) {\n\t\t\tshowPreviewUI(next);\n\t\t} else {\n\t\t\thidePreviewUI();\n\t\t}\n\t} catch (err) {\n\t\thidePreviewUI();\n\t}\n}\n\n// -------------------- SGLang Client (simple test) --------------------\nasync function callSGLangChat(): Promise<void> {\n\tconst config = vscode.workspace.getConfiguration();\n\t\n\tlet hostname: string;\n\tlet port: number;\n\tlet path: string;\n\tlet useHttps = true;\n\tlet modelName: string;\n\tconst headers: any = {\n\t\t'Content-Type': 'application/json'\n\t};\n\n\tif (!USE_GEMINI) {\n\t\t// SGLang\n\t\thostname = SGLANG_HOSTNAME;\n\t\tport = SGLANG_PORT;\n\t\tpath = SGLANG_BASE_PATH;\n\t\tuseHttps = false; \n\t\tmodelName = SGLANG_MODEL_NAME;\n\t} else {\n\t\t// Gemini\n\t\tconst apiKey = config.get<string>('crowd-pilot.apiKey');\n\t\tif (!apiKey) {\n\t\t\tvscode.window.showErrorMessage('Crowd Pilot: Please set your API Key in settings (crowd-pilot.apiKey).');\n\t\t\treturn;\n\t\t}\n\t\thostname = GEMINI_HOSTNAME;\n\t\tport = GEMINI_PORT;\n\t\tpath = GEMINI_BASE_PATH;\n\t\theaders['Authorization'] = `Bearer ${apiKey}`;\n\t\tmodelName = GEMINI_MODEL_NAME;\n\t}\n\n\tconst requestBody = {\n\t\tmodel: modelName,\n\t\tmessages: [\n\t\t\t{ role: 'user', content: 'What is the capital of France?' }\n\t\t]\n\t};\n\tconst postData = JSON.stringify(requestBody);\n\theaders['Content-Length'] = Buffer.byteLength(postData);\n\n\tconst options = {\n\t\thostname,\n\t\tport,\n\t\tpath,\n\t\tmethod: 'POST',\n\t\theaders\n\t};\n\n\tconst requestModule = useHttps ? https : http;\n\n\ttry {\n\t\tconst json = await new Promise<any>((resolve, reject) => {\n\t\t\tconst req = requestModule.request(options, (res: http.IncomingMessage) => {\n\t\t\t\tlet data = '';\n\t\t\t\tres.on('data', (chunk: Buffer) => {\n\t\t\t\t\tdata += chunk.toString();\n\t\t\t\t});\n\t\t\t\tres.on('end', () => {\n\t\t\t\t\ttry {\n\t\t\t\t\t\tresolve(JSON.parse(data));\n\t\t\t\t\t} catch (err) {\n\t\t\t\t\t\treject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n\t\t\t\t\t}\n\t\t\t\t});\n\t\t\t});\n\n\t\t\treq.on('error', (err: Error) => {\n\t\t\t\treject(err);\n\t\t\t});\n\n\t\t\treq.write(postData);\n\t\t\treq.end();\n\t\t});\n\n\t\tvscode.window.showInformationMessage(`Response: ${JSON.stringify(json, null, 2)}`);\n\t} catch (err) {\n\t\tconst errorMessage = err instanceof Error ? err.message : String(err);\n\t\tvscode.window.showErrorMessage(`Request failed: ${errorMessage}`);\n\t}\n}\n\n// -------------------- Model-planned Actions --------------------\nasync function requestModelActions(editor: vscode.TextEditor): Promise<PlannedAction> {\n\tconst config = vscode.workspace.getConfiguration();\n\t\n\tlet hostname: string;\n\tlet port: number;\n\tlet path: string;\n\tlet useHttps = true;\n\tlet modelName: string;\n\tconst headers: any = {\n\t\t'Content-Type': 'application/json'\n\t};\n\n\tif (!USE_GEMINI) {\n\t\t// SGLang\n\t\thostname = SGLANG_HOSTNAME;\n\t\tport = SGLANG_PORT;\n\t\tpath = SGLANG_BASE_PATH;\n\t\tuseHttps = false;\n\t\tmodelName = SGLANG_MODEL_NAME;\n\t} else {\n\t\t// Gemini\n\t\tconst apiKey = config.get<string>('crowd-pilot.apiKey');\n\t\tif (!apiKey) {\n\t\t\tvscode.window.showErrorMessage('Crowd Pilot: Please set your API Key in settings (crowd-pilot.apiKey).');\n\t\t\tthrow new Error('API key not set');\n\t\t}\n\t\thostname = GEMINI_HOSTNAME;\n\t\tport = GEMINI_PORT;\n\t\tpath = GEMINI_BASE_PATH;\n\t\theaders['Authorization'] = `Bearer ${apiKey}`;\n\t\tmodelName = GEMINI_MODEL_NAME;\n\t}\n\n\tconst schemaDescription = [\n\t\t'Role: You suggest the next VS Code editor/terminal action to progress the current task.',\n\t\t'Output ONLY a JSON object (no prose, no code fences).',\n\t\t'Coordinates are zero-based [line, column].',\n\t\t'Allowed actions (JSON schema-like):',\n\t\t'{ kind: ""showTextDocument"" }',\n\t\t'{ kind: ""setSelections"", selections: Array<{ start: [number, number], end: [number, number] }> }',\n\t\t'{ kind: ""editInsert"", position: [number, number], text: string }',\n\t\t'{ kind: ""editDelete"", range: { start: [number, number], end: [number, number] } }',\n\t\t'{ kind: ""editReplace"", range: { start: [number, number], end: [number, number] }, text: string }',\n\t\t'{ kind: ""terminalShow"" }',\n\t\t'{ kind: ""terminalSendText"", text: string }',\n\t\t'Guidelines:',\n\t\t'- If you you insert text, insert until the logical end of the current statement or block.',\n\t\t'- When inserting text, make sure to not repeat existing text (except when replacing existing text).',\n\t\t'- Use double-quoted JSON strings.'\n\t].join('\n');\n\n\tconst doc = editor.document;\n\tconst cursor = editor.selection.active;\n\tconst fullText = doc.getText();\n\tconst numberedContext = fullText.split(/\r?\n/).map((line, i) => `${i}: ${line}`).join('\n');\n\n\tconst tabbingPrompt = [\n\t\t'Your role: Propose the single next action according to the schema to help the developer progress.',\n\t\t'',\n\t\t'Available context:',\n\t\t`- File: ${doc.fileName}`,\n\t\t`- Language: ${doc.languageId}`,\n\t\t`- Cursor: (${cursor.line}, ${cursor.character})`,\n\t\t'',\n\t\t'Full file content (zero-based line numbers):',\n\t\t'```',\n\t\tnumberedContext,\n\t\t'```',\n\t\t'',\n\t\t'Respond with ONLY a JSON object containing exactly one action.'\n\t].join('\n');\n\n\tconst requestBody = {\n\t\tmodel: modelName,\n\t\tmessages: [\n\t\t\t{ role: 'system', content: schemaDescription },\n\t\t\t{ role: 'user', content: tabbingPrompt }\n\t\t]\n\t};\n\n\tconst postData = JSON.stringify(requestBody);\n\theaders['Content-Length'] = Buffer.byteLength(postData);\n\n\tconst options = {\n\t\thostname,\n\t\tport,\n\t\tpath,\n\t\tmethod: 'POST',\n\t\theaders\n\t};\n\n\tconst requestModule = useHttps ? https : http;\n\n\tconst json = await new Promise<any>((resolve, reject) => {\n\t\tconst req = requestModule.request(options, (res: http.IncomingMessage) => {\n",typescript,content
+5,5992314,"/fast/home/franz.srambical/crowd-pilot-extension/out/extension.js",0,0,"""use strict"";\nvar __createBinding = (this && this.__createBinding) || (Object.create ? (function(o, m, k, k2) {\n    if (k2 === undefined) k2 = k;\n    var desc = Object.getOwnPropertyDescriptor(m, k);\n    if (!desc || (""get"" in desc ? !m.__esModule : desc.writable || desc.configurable)) {\n      desc = { enumerable: true, get: function() { return m[k]; } };\n    }\n    Object.defineProperty(o, k2, desc);\n}) : (function(o, m, k, k2) {\n    if (k2 === undefined) k2 = k;\n    o[k2] = m[k];\n}));\nvar __setModuleDefault = (this && this.__setModuleDefault) || (Object.create ? (function(o, v) {\n    Object.defineProperty(o, ""default"", { enumerable: true, value: v });\n}) : function(o, v) {\n    o[""default""] = v;\n});\nvar __importStar = (this && this.__importStar) || (function () {\n    var ownKeys = function(o) {\n        ownKeys = Object.getOwnPropertyNames || function (o) {\n            var ar = [];\n            for (var k in o) if (Object.prototype.hasOwnProperty.call(o, k)) ar[ar.length] = k;\n            return ar;\n        };\n        return ownKeys(o);\n    };\n    return function (mod) {\n        if (mod && mod.__esModule) return mod;\n        var result = {};\n        if (mod != null) for (var k = ownKeys(mod), i = 0; i < k.length; i++) if (k[i] !== ""default"") __createBinding(result, mod, k[i]);\n        __setModuleDefault(result, mod);\n        return result;\n    };\n})();\nObject.defineProperty(exports, ""__esModule"", { value: true });\nexports.activate = activate;\nexports.deactivate = deactivate;\nconst vscode = __importStar(require(""vscode""));\nconst http = __importStar(require(""http""));\nconst buffer_1 = require(""buffer"");\nconst HOSTNAME = 'hai001';\nconst PORT = 30000;\nfunction activate(context) {\n    console.log('[crowd-pilot] Extension activated');\n    // Configure terminal to allow tab keybinding to work\n    (async () => {\n        const config = vscode.workspace.getConfiguration('terminal.integrated');\n        const commandsToSkipShell = config.get('commandsToSkipShell', []);\n        let updated = false;\n        if (!commandsToSkipShell.includes('crowd-pilot.modelRun')) {\n            commandsToSkipShell.push('crowd-pilot.modelRun');\n            updated = true;\n        }\n        if (!commandsToSkipShell.includes('crowd-pilot.hideUi')) {\n            commandsToSkipShell.push('crowd-pilot.hideUi');\n            updated = true;\n        }\n        if (updated) {\n            await config.update('commandsToSkipShell', commandsToSkipShell, vscode.ConfigurationTarget.Global);\n        }\n    })().catch((err) => console.error('[crowd-pilot] Startup initialization error:', err));\n    const hideUi = vscode.commands.registerCommand('crowd-pilot.hideUi', () => {\n        hidePreviewUI(true);\n    });\n    const modelRun = vscode.commands.registerCommand('crowd-pilot.modelRun', async () => {\n        const editor = vscode.window.activeTextEditor;\n        if (!editor) {\n            return;\n        }\n        try {\n            // Confirm only when a suggestion is visible\n            if (!previewVisible) {\n                return;\n            }\n            let action = currentAction;\n            if (!action) {\n                const single = await requestModelActions(editor);\n                currentAction = single;\n                action = single;\n            }\n            if (!action) {\n                hidePreviewUI();\n                return;\n            }\n            hidePreviewUI(false);\n            await executeAction(action);\n            autoShowNextAction();\n        }\n        catch (err) {\n            const errorMessage = err instanceof Error ? err.message : String(err);\n            vscode.window.showErrorMessage(`Model run failed: ${errorMessage}`);\n        }\n    });\n    const sglangTest = vscode.commands.registerCommand('crowd-pilot.sglangTest', async () => {\n        try {\n            await callSGLangChat();\n        }\n        catch (err) {\n            const errorMessage = err instanceof Error ? err.message : String(err);\n            vscode.window.showErrorMessage(`SGLang test failed: ${errorMessage}`);\n        }\n    });\n    // Auto-preview listeners\n    const onSelChange = vscode.window.onDidChangeTextEditorSelection((e) => {\n        if (e.textEditor === vscode.window.activeTextEditor) {\n            suppressAutoPreview = false;\n            autoShowNextAction();\n        }\n    });\n    const onActiveChange = vscode.window.onDidChangeActiveTextEditor(() => {\n        suppressAutoPreview = false;\n        autoShowNextAction();\n    });\n    const onDocChange = vscode.workspace.onDidChangeTextDocument((e) => {\n        if (vscode.window.activeTextEditor?.document === e.document) {\n            suppressAutoPreview = false;\n            autoShowNextAction();\n        }\n    });\n    context.subscriptions.push(hideUi, sglangTest, modelRun, onSelChange, onActiveChange, onDocChange);\n}\nfunction deactivate() { }\nlet currentAction;\nasync function executeAction(action) {\n    const editor = vscode.window.activeTextEditor;\n    if (!editor) {\n        return;\n    }\n    const doc = editor.document;\n    const term = vscode.window.terminals[0] ?? vscode.window.createTerminal('Test');\n    if (action.kind === 'showTextDocument') {\n        await vscode.window.showTextDocument(doc);\n        return;\n    }\n    if (action.kind === 'setSelections') {\n        editor.selections = action.selections.map(s => new vscode.Selection(new vscode.Position(s.start[0], s.start[1]), new vscode.Position(s.end[0], s.end[1])));\n        if (editor.selections.length > 0) {\n            editor.revealRange(editor.selections[0], vscode.TextEditorRevealType.InCenterIfOutsideViewport);\n        }\n        return;\n    }\n    if (action.kind === 'editInsert') {\n        await editor.edit((e) => e.insert(new vscode.Position(action.position[0], action.position[1]), action.text));\n        return;\n    }\n    if (action.kind === 'editDelete') {\n        const range = new vscode.Range(new vscode.Position(action.range.start[0], action.range.start[1]), new vscode.Position(action.range.end[0], action.range.end[1]));\n        await editor.edit((e) => e.delete(range));\n        return;\n    }\n    if (action.kind === 'editReplace') {\n        const range = new vscode.Range(new vscode.Position(action.range.start[0], action.range.start[1]), new vscode.Position(action.range.end[0], action.range.end[1]));\n        await editor.edit((e) => e.replace(range, action.text));\n        return;\n    }\n    if (action.kind === 'terminalShow') {\n        term.show();\n        return;\n    }\n    if (action.kind === 'terminalSendText') {\n        term.sendText(action.text);\n        return;\n    }\n}\n// -------------------- UI State & Helpers --------------------\nconst UI_CONTEXT_KEY = 'crowdPilot.uiVisible';\nlet previewVisible = false;\nlet decorationDeleteType;\nlet decorationReplaceType;\nlet decorationReplaceBlockType;\nlet mockStep = 0;\nlet suppressAutoPreview = false;\nfunction disposePreviewDecorations() {\n    try {\n        decorationDeleteType?.dispose();\n    }\n    catch { }\n    try {\n        decorationReplaceType?.dispose();\n    }\n    catch { }\n    try {\n        decorationReplaceBlockType?.dispose();\n    }\n    catch { }\n    decorationDeleteType = undefined;\n    decorationReplaceType = undefined;\n    decorationReplaceBlockType = undefined;\n}\nfunction getDynamicMargin(editor, anchorLine, text) {\n    // Count lines in the preview text\n    const lines = text.split(/\r?\n/);\n    const height = lines.length;\n    // We need to check the document lines that will be covered by this panel.\n    // The panel starts at 'anchorLine' and extends downwards by 'height' lines.\n    // However, visually, since it's 'after', it sits to the right of 'anchorLine',\n    // and then flows down.\n    // So we check document lines from anchorLine to anchorLine + height - 1.\n    const doc = editor.document;\n    let maxLen = 0;\n    const startLine = anchorLine;\n    const endLine = Math.min(doc.lineCount - 1, anchorLine + height - 1);\n    for (let i = startLine; i <= endLine; i++) {\n        const lineText = doc.lineAt(i).text;\n        // Simple approximation: assume tabs are 4 spaces if we can't get config easily, \n        // or just treat them as 1 char (which might underestimate). \n        // Better to overestimate: treat tab as 4 chars.\n        const len = lineText.replace(/\t/g, '    ').length;\n        if (len > maxLen) {\n            maxLen = len;\n        }\n    }\n    // Length of the anchor line itself\n    const anchorLineText = doc.lineAt(anchorLine).text;\n    const anchorLen = anchorLineText.replace(/\t/g, '    ').length;\n    // The offset needed is maxLen - anchorLen.\n    // If maxLen <= anchorLen, offset is 0 (margin is just base padding).\n    // If maxLen > anchorLen, we need to push right by (maxLen - anchorLen).\n    const diff = Math.max(0, maxLen - anchorLen);\n    // Base margin 2rem is roughly 4ch. Let's use ch units for everything to be consistent.\n    // 1ch is width of '0'. In monospace, mostly consistent.\n    // Add 3ch extra padding for safety/visual gap.\n    const margin = diff + 4;\n    return `${margin}ch`;\n}\nfunction showPreviewUI(action) {\n    const editor = vscode.window.activeTextEditor;\n    if (!editor) {\n        return;\n    }\n    disposePreviewDecorations();\n    // FIXME (f.srambical): add file switch \n    const next = (action.kind === 'editInsert' || action.kind === 'editDelete' || action.kind === 'editReplace' || action.kind === 'terminalSendText' || action.kind === 'setSelections') ? action : undefined;\n    if (!next) {\n        previewVisible = false;\n        vscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n        currentAction = action;\n        return;\n    }\n    const trimText = (t) => {\n        const oneLine = t.replace(/\r?\n/g, '\\n');\n        return oneLine.length > 80 ? oneLine.slice(0, 77) + '…' : oneLine;\n    };\n    if (next.kind === 'setSelections') {\n        // For setSelections, we only preview the primary selection's start/active position\n        const selection = next.selections[0];\n        const targetPos = new vscode.Position(selection.start[0], selection.start[1]);\n        // Check if the target position is visible\n        const isVisible = editor.visibleRanges.some(r => r.contains(targetPos));\n        let anchorPos = targetPos;\n        let label = ""↳ Move Cursor Here"";\n        if (!isVisible && editor.visibleRanges.length > 0) {\n            const firstVisible = editor.visibleRanges[0].start;\n            const lastVisible = editor.visibleRanges[editor.visibleRanges.length - 1].end;\n            if (targetPos.isBefore(firstVisible)) {\n                anchorPos = editor.document.lineAt(firstVisible.line).range.end;\n            }\n            else {\n                anchorPos = editor.document.lineAt(lastVisible.line).range.end;\n            }\n            if (targetPos.line < anchorPos.line) {\n                label = `↑ Move Cursor to Line ${targetPos.line + 1}`;\n            }\n            else {\n                label = `↓ Move Cursor to Line ${targetPos.line + 1}`;\n            }\n        }\n        const margin = getDynamicMargin(editor, anchorPos.line, label);\n        decorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n            after: {\n                contentText: '',\n                color: new vscode.ThemeColor('charts.purple'),\n                backgroundColor: new vscode.ThemeColor('editor.background'),\n                fontStyle: 'italic',\n                fontWeight: '600',\n                margin: `0 0 0 ${margin}`,\n                textDecoration: `none; display: inline-block; white-space: pre; content: ""${label}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n            }\n        });\n        editor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n    }\n    else if (next.kind === 'terminalSendText') {\n        const cursor = editor.selection.active;\n        const lineEnd = editor.document.lineAt(cursor.line).range.end;\n        const cmd = next.text.replace(/""/g, '\\""').replace(/\r?\n/g, '\\A ');\n        const margin = getDynamicMargin(editor, cursor.line, ""↳ Execute in Terminal:\n"" + next.text);\n        decorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n            after: {\n                contentText: '',\n                color: new vscode.ThemeColor('charts.purple'),\n                backgroundColor: new vscode.ThemeColor('editor.background'),\n                fontStyle: 'italic',\n                fontWeight: '600',\n                margin: `0 0 0 ${margin}`,\n                textDecoration: `none; display: inline-block; white-space: pre; content: ""↳ Execute in Terminal:\\A ${cmd}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n            }\n        });\n        editor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(lineEnd, lineEnd) }]);\n    }\n    else if (next.kind === 'editInsert') {\n        const posLine = next.position[0];\n        const fullBlock = next.text;\n        const cssContent = fullBlock\n            .replace(/""/g, '\\""')\n            .replace(/\r?\n/g, '\\A ');\n        const docLineCount = editor.document.lineCount;\n        // If inserting at EOF (or beyond), attach to the last line.\n        // Otherwise, attach to the line AT the insertion point and shift visually UP into the gap.\n        let anchorLine = posLine;\n        let shiftUp = true;\n        if (anchorLine >= docLineCount) {\n            anchorLine = docLineCount - 1;\n            shiftUp = false; // At EOF, we just append below or to the right\n        }\n        const anchorPos = new vscode.Position(anchorLine, Number.MAX_VALUE);\n        // We attach to the line AT the insertion point.\n        // The panel floats to the right of this line.\n        // The dashed line connects the start of this line to the panel.\n        // This indicates that the new text will be inserted at this line position (pushing the current line down).\n        const marginCheckLine = anchorLine;\n        const margin = getDynamicMargin(editor, marginCheckLine, fullBlock);\n        const topOffset = '0';\n        // Dashed line style\n        // We use 'before' decoration for the line.\n        // It needs to be absolute, full width (or enough to reach left), \n        // and aligned with the panel top.\n        const beforeDecoration = {\n            contentText: '',\n            textDecoration: `none; position: absolute; left: 0; width: 100vw; border-top: 1px dashed var(--vscode-charts-purple); top: 0; height: 0; z-index: 99; pointer-events: none;`\n        };\n        decorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n            before: beforeDecoration,\n            after: {\n                contentText: '',\n                color: new vscode.ThemeColor('charts.purple'),\n                backgroundColor: new vscode.ThemeColor('editor.background'),\n                fontStyle: 'italic',\n                fontWeight: '600',\n                margin: `0 0 0 ${margin}`,\n                textDecoration: `none; display: inline-block; white-space: pre; content: ""${cssContent}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top; top: ${topOffset};`\n            }\n        });\n        editor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n    }\n    else if (next.kind === 'editDelete') {\n        const range = new vscode.Range(new vscode.Position(next.range.start[0], next.range.start[1]), new vscode.Position(next.range.end[0], next.range.end[1]));\n        decorationDeleteType = vscode.window.createTextEditorDecorationType({\n            backgroundColor: 'rgba(255, 60, 60, 0.18)',\n            border: '1px solid rgba(255, 60, 60, 0.35)',\n            textDecoration: 'line-through'\n        });\n        editor.setDecorations(decorationDeleteType, [{ range }]);\n    }\n    else if (next.kind === 'editReplace') {\n        const range = new vscode.Range(new vscode.Position(next.range.start[0], next.range.start[1]), new vscode.Position(next.range.end[0], next.range.end[1]));\n        // Highlight original range (to be replaced)\n        decorationReplaceType = vscode.window.createTextEditorDecorationType({\n            backgroundColor: 'rgba(255,165,0,0.15)',\n            border: '1px dashed rgba(255,165,0,0.45)',\n            color: new vscode.ThemeColor('disabledForeground'),\n            textDecoration: 'line-through'\n        });\n        editor.setDecorations(decorationReplaceType, [{ range }]);\n        // Show replacement block to the right of the first replaced line\n        const fullBlock = next.text;\n        // CSS-escape the text for the 'content' property:\n        // - Escape double quotes\n        // - Replace newlines with \A (CSS newline)\n        const cssContent = fullBlock\n            .replace(/""/g, '\\""')\n            .replace(/\r?\n/g, '\\A ');\n        // Attach 'after' decoration to the start of the replacement range\n        // (Actually, attaching to the end of the first line is safer for 'after')\n        const anchorLine = range.start.line;\n        const anchorPos = new vscode.Position(anchorLine, Number.MAX_VALUE);\n        const margin = getDynamicMargin(editor, anchorLine, fullBlock);\n        decorationReplaceBlockType = vscode.window.createTextEditorDecorationType({\n            after: {\n                contentText: '', // Handled by CSS content\n                color: new vscode.ThemeColor('charts.purple'),\n                backgroundColor: new vscode.ThemeColor('editor.background'),\n                fontStyle: 'italic',\n                fontWeight: '600',\n                margin: `0 0 0 ${margin}`,\n                textDecoration: `none; display: inline-block; white-space: pre; content: ""${cssContent}""; border: 1px solid var(--vscode-charts-purple); padding: 4px; border-radius: 4px; box-shadow: 0 4px 8px rgba(0,0,0,0.25); pointer-events: none; position: relative; z-index: 100; vertical-align: top;`\n            }\n        });\n        editor.setDecorations(decorationReplaceBlockType, [{ range: new vscode.Range(anchorPos, anchorPos) }]);\n    }\n    previewVisible = true;\n    vscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, true);\n    currentAction = action;\n}\nfunction hidePreviewUI(suppress) {\n    disposePreviewDecorations();\n    previewVisible = false;\n    vscode.commands.executeCommand('setContext', UI_CONTEXT_KEY, false);\n    if (suppress) {\n        suppressAutoPreview = true;\n    }\n}\n// -------------------- Hardcoded single-step actions --------------------\nfunction getHardcodedNextAction(editor) {\n    const cursor = editor.selection.active;\n    const doc = editor.document;\n    const lineCount = doc.lineCount;\n    const clamp = (n, min, max) => Math.max(min, Math.min(max, n));\n    // Step 0: Insert multiline content two lines below the cursor (start of target line)\n    if (mockStep === 0) {\n        const targetLine = clamp(cursor.line + 2, 0, Math.max(0, lineCount - 1));\n        return {\n            kind: 'editInsert',\n            position: [targetLine, 0],\n            text: '/* crowd-pilot: insert start */\nline A\nline B\n/* crowd-pilot: insert end */\n'\n        };\n    }\n    // Step 1: Replace a two-line range three and four lines below the cursor\n    if (mockStep === 1) {\n        const startLine = clamp(cursor.line + 3, 0, Math.max(0, lineCount - 1));\n        const endLine = clamp(startLine + 1, 0, Math.max(0, lineCount - 1));\n        const endChar = doc.lineAt(endLine).range.end.character;\n        const range = {\n            start: [startLine, 0],\n            end: [endLine, endChar]\n        };\n        const replacement = [\n            '/* crowd-pilot: replacement */',\n            'REPLACED LINE 1',\n            'REPLACED LINE 2'\n        ].join('\n');\n        return { kind: 'editReplace', range, text: replacement };\n    }\n    // Step 2: Delete a three-line range six to eight lines below the cursor\n    if (mockStep === 2) {\n        const startLine = clamp(cursor.line + 6, 0, Math.max(0, lineCount - 1));\n        const endLine = clamp(startLine + 2, 0, Math.max(0, lineCount - 1));\n        // To fully delete the lines including the newline, we target the start of the next line.\n        let endPosLine = endLine + 1;\n        let endPosChar = 0;\n        if (endPosLine >= lineCount) {\n            // If deleting the last line(s), just go to the end of the document\n            endPosLine = lineCount - 1;\n            endPosChar = doc.lineAt(endPosLine).range.end.character;\n        }\n        const range = {\n            start: [startLine, 0],\n            end: [endPosLine, endPosChar]\n        };\n        return { kind: 'editDelete', range };\n    }\n    // Step 3: Execute in Terminal\n    if (mockStep === 3) {\n        return { kind: 'terminalSendText', text: 'echo ""Hello World""' };\n    }\n    // Step 4: Move Cursor to End of File\n    if (mockStep === 4) {\n        const lastLine = doc.lineCount - 1;\n        const lastChar = doc.lineAt(lastLine).range.end.character;\n        return {\n            kind: 'setSelections',\n            selections: [{ start: [lastLine, lastChar], end: [lastLine, lastChar] }]\n        };\n    }\n    return undefined;\n}\nfunction advanceMockStep() {\n    mockStep = (mockStep + 1) % 5;\n}\nasync function autoShowNextAction() {\n    if (suppressAutoPreview) {\n        return;\n    }\n    const editor = vscode.window.activeTextEditor;\n    if (!editor) {\n        return;\n    }\n    try {\n        const next = await requestModelActions(editor);\n        if (next) {\n            showPreviewUI(next);\n        }\n        else {\n            hidePreviewUI();\n        }\n    }\n    catch (err) {\n        hidePreviewUI();\n    }\n}\n// -------------------- SGLang Client (simple test) --------------------\nasync function callSGLangChat() {\n    const requestBody = {\n        model: 'qwen/qwen2.5-0.5b-instruct',\n        messages: [\n            { role: 'user', content: 'What is the capital of France?' }\n        ]\n    };\n    const postData = JSON.stringify(requestBody);\n    const options = {\n        hostname: HOSTNAME,\n        port: PORT,\n        path: '/v1/chat/completions',\n        method: 'POST',\n        headers: {\n            'Content-Type': 'application/json',\n            'Content-Length': buffer_1.Buffer.byteLength(postData)\n        }\n    };\n    try {\n        const json = await new Promise((resolve, reject) => {\n            const req = http.request(options, (res) => {\n                let data = '';\n                res.on('data', (chunk) => {\n                    data += chunk.toString();\n                });\n                res.on('end', () => {\n                    try {\n                        resolve(JSON.parse(data));\n                    }\n                    catch (err) {\n                        reject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n                    }\n                });\n            });\n            req.on('error', (err) => {\n                reject(err);\n            });\n            req.write(postData);\n            req.end();\n        });\n        vscode.window.showInformationMessage(`SGLang response: ${JSON.stringify(json, null, 2)}`);\n    }\n    catch (err) {\n        const errorMessage = err instanceof Error ? err.message : String(err);\n        vscode.window.showErrorMessage(`SGLang request failed: ${errorMessage}`);\n    }\n}\n// -------------------- Model-planned Actions --------------------\nasync function requestModelActions(editor) {\n    const schemaDescription = [\n        'Role: You suggest the next VS Code editor/terminal action to progress the current task.',\n        'Output ONLY a JSON object (no prose, no code fences).',\n        'Coordinates are zero-based [line, column].',\n        'Allowed actions (JSON schema-like):',\n        '{ kind: ""showTextDocument"" }',\n        '{ kind: ""setSelections"", selections: Array<{ start: [number, number], end: [number, number] }> }',\n        '{ kind: ""editInsert"", position: [number, number], text: string }',\n        '{ kind: ""editDelete"", range: { start: [number, number], end: [number, number] } }',\n        '{ kind: ""editReplace"", range: { start: [number, number], end: [number, number] }, text: string }',\n        '{ kind: ""terminalShow"" }',\n        '{ kind: ""terminalSendText"", text: string }',\n        'Guidelines:',\n        '- If you you insert text, insert until the logical end of the current statement or block.',\n        '- When inserting text, make sure to not repeat existing text (except when replacing existing text).',\n        '- Use double-quoted JSON strings.'\n    ].join('\n');\n    const doc = editor.document;\n    const cursor = editor.selection.active;\n    const fullText = doc.getText();\n    const numberedContext = fullText.split(/\r?\n/).map((line, i) => `${i}: ${line}`).join('\n');\n    const tabbingPrompt = [\n        'Your role: Propose the single next action according to the schema to help the developer progress.',\n        '',\n        'Available context:',\n        `- File: ${doc.fileName}`,\n        `- Language: ${doc.languageId}`,\n        `- Cursor: (${cursor.line}, ${cursor.character})`,\n        '',\n        'Full file content (zero-based line numbers):',\n        '```',\n        numberedContext,\n        '```',\n        '',\n        'Respond with ONLY a JSON object containing exactly one action.'\n    ].join('\n');\n    const requestBody = {\n        model: 'qwen/qwen2.5-0.5b-instruct',\n        messages: [\n            { role: 'system', content: schemaDescription },\n            { role: 'user', content: tabbingPrompt }\n        ]\n    };\n    const postData = JSON.stringify(requestBody);\n    const options = {\n        hostname: HOSTNAME,\n        port: PORT,\n        path: '/v1/chat/completions',\n        method: 'POST',\n        headers: {\n            'Content-Type': 'application/json',\n            'Content-Length': buffer_1.Buffer.byteLength(postData)\n        }\n    };\n    const json = await new Promise((resolve, reject) => {\n        const req = http.request(options, (res) => {\n            let data = '';\n            res.on('data', (chunk) => { data += chunk.toString(); });\n            res.on('end', () => {\n                try {\n                    resolve(JSON.parse(data));\n                }\n                catch (err) {\n                    reject(new Error(`Failed to parse response: ${err instanceof Error ? err.message : String(err)}`));\n                }\n            });\n        });\n        req.on('error', (err) => reject(err));\n        req.write(postData);\n        req.end();\n    });\n    const content = extractChatContent(json);\n    if (typeof content !== 'string' || content.trim().length === 0) {\n        throw new Error('Empty model content');\n    }\n    const action = parsePlannedAction(content);\n    if (!action) {\n        throw new Error('No valid action parsed from model output');\n    }\n    return action;\n}\nfunction extractChatContent(json) {\n    try {\n        if (json && Array.isArray(json.choices) && json.choices[0]) {\n            const choice = json.choices[0];\n            if (choice.message && typeof choice.message.content === 'string') {\n                return choice.message.content;\n            }\n            if (typeof choice.text === 'string') {\n                return choice.text;\n            }\n        }\n        return undefined;\n    }\n    catch {\n        return undefined;\n    }\n}\nfunction parsePlannedAction(raw) {\n    let text = raw.trim();\n    text = text.replace(/^```(?:json)?\s*/i, '').replace(/```\s*$/i, '').trim();\n    text = text.replace(/<think>[\s\S]*?<\/think>/gi, '').trim();\n    let parsed;\n    try {\n        parsed = JSON.parse(text);\n    }\n    catch (err) {\n        return undefined;\n    }\n    if (Array.isArray(parsed)) {\n        console.error('Model should not return an array.');\n        return undefined;\n    }\n    switch (parsed.kind) {\n        case 'showTextDocument':\n            return { kind: 'showTextDocument' };\n        case 'setSelections': {\n            const selections = Array.isArray(parsed.selections) ? parsed.selections : [];\n            const norm = selections.map((s) => ({\n                start: Array.isArray(s?.start) && s.start.length === 2 ? [Number(s.start[0]) || 0, Number(s.start[1]) || 0] : [0, 0],\n                end: Array.isArray(s?.end) && s.end.length === 2 ? [Number(s.end[0]) || 0, Number(s.end[1]) || 0] : [0, 0]\n            }));\n            return { kind: 'setSelections', selections: norm };\n        }\n        case 'editInsert': {\n            const pos = Array.isArray(parsed.position) && parsed.position.length === 2 ? [Number(parsed.position[0]) || 0, Number(parsed.position[1]) || 0] : [0, 0];\n            const textVal = typeof parsed.text === 'string' ? parsed.text : '';\n            return { kind: 'editInsert', position: pos, text: textVal };\n        }\n        case 'editDelete': {\n            const start = Array.isArray(parsed.range?.start) && parsed.range.start.length === 2 ? [Number(parsed.range.start[0]) || 0, Number(parsed.range.start[1]) || 0] : [0, 0];\n            const end = Array.isArray(parsed.range?.end) && parsed.range.end.length === 2 ? [Number(parsed.range.end[0]) || 0, Number(parsed.range.end[1]) || 0] : [0, 0];\n            return { kind: 'editDelete', range: { start, end } };\n        }\n        case 'editReplace': {\n            const start = Array.isArray(parsed.range?.start) && parsed.range.start.length === 2 ? [Number(parsed.range.start[0]) || 0, Number(parsed.range.start[1]) || 0] : [0, 0];\n            const end = Array.isArray(parsed.range?.end) && parsed.range.end.length === 2 ? [Number(parsed.range.end[0]) || 0, Number(parsed.range.end[1]) || 0] : [0, 0];\n            const textVal = typeof parsed.text === 'string' ? parsed.text : '';\n            return { kind: 'editReplace', range: { start, end }, text: textVal };\n        }\n        case 'terminalShow':\n            return { kind: 'terminalShow' };\n        case 'terminalSendText': {\n            const textVal = typeof parsed.text === 'string' ? parsed.text : '';\n            return { kind: 'terminalSendText', text: textVal };\n        }\n        default:\n            return undefined;\n    }\n}\n//# sourceMappingURL=extension.js.map",javascript,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-70e1b9a4-9447-4be7-85f9-ea3b825230fe1765227767400-2025_12_08-22.02.58.450/source.csv

@@ -0,0 +1,103 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,558,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:02:58 PM [info] Activating crowd-code\n10:02:58 PM [info] Recording started\n10:02:58 PM [info] Initializing git provider using file system watchers...\n10:02:58 PM [info] No workspace folder found\n",Log,tab
+3,2279,"extension-output-pdoom-org.crowd-code-#1-crowd-code",198,0,"10:03:00 PM [info] Retrying git provider initialization...\n10:03:00 PM [info] No workspace folder found\n",Log,content
+4,419530,"Untitled-1",0,0,"",plaintext,tab
+5,424918,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",0,0,"#!/usr/bin/env python3\nimport sys\nfrom pathlib import Path\nimport pandas as pd\n\ninput_dir = sys.argv[1]\nfor parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df['text']:\n        print(text)\n",python,tab
+6,426336,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"\n",python,content
+7,452818,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,0,"import os\\nif not os.path.exists(input_dir):\\n    print(f""Error: Input directory {input_dir} does not exist."")\\n    sys.exit(1)\n",python,content
+8,467833,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",388,0,"",python,selection_command
+9,467989,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",368,0,"",python,selection_command
+10,468134,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",340,0,"",python,selection_command
+11,468281,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",301,0,"",python,selection_command
+12,468429,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",232,0,"",python,selection_command
+13,468617,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",216,0,"",python,selection_command
+14,469181,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",216,15,"    sys.exit(1)",python,selection_command
+15,469520,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",150,81,"    print(f""Error: Input directory {input_dir} does not exist."")\\n    sys.exit(1)",python,selection_command
+16,469692,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",115,116,"if not os.path.exists(input_dir):\\n    print(f""Error: Input directory {input_dir} does not exist."")\\n    sys.exit(1)",python,selection_command
+17,469885,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",115,117,"",python,content
+18,471492,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,0,"",python,selection_command
+19,471841,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,11,"",python,content
+20,472894,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",173,0,"",python,selection_command
+21,473130,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",212,0,"",python,selection_command
+22,473169,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",240,0,"",python,selection_command
+23,473198,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+24,473228,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"",python,selection_command
+25,476710,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"d",python,content
+26,476712,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"",python,selection_keyboard
+27,476815,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"e",python,content
+28,476818,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"",python,selection_keyboard
+29,477038,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"f",python,content
+30,477041,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"",python,selection_keyboard
+31,477144,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0," ",python,content
+32,477147,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_keyboard
+33,478878,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,1,"",python,content
+34,479049,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,1,"",python,content
+35,479205,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,1,"",python,content
+36,479362,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,1,"",python,content
+37,481961,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"i",python,content
+38,481963,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"",python,selection_keyboard
+39,482194,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"f",python,content
+40,482196,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"",python,selection_keyboard
+41,482236,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0," ",python,content
+42,482239,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"",python,selection_keyboard
+43,484322,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"_",python,content
+44,484325,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_keyboard
+45,484478,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"_",python,content
+46,484481,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"",python,selection_keyboard
+47,484810,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"n",python,content
+48,484813,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"",python,selection_keyboard
+49,484990,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"a",python,content
+50,484992,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"",python,selection_keyboard
+51,485114,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"m",python,content
+52,485116,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"",python,selection_keyboard
+53,485307,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"e",python,content
+54,485309,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"",python,selection_keyboard
+55,485725,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"_",python,content
+56,485728,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"",python,selection_keyboard
+57,485948,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"_",python,content
+58,485950,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0,"",python,selection_keyboard
+59,486643,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0," ",python,content
+60,486645,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"",python,selection_keyboard
+61,487699,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"=",python,content
+62,487702,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0,"",python,selection_keyboard
+63,487757,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0,"=",python,content
+64,487760,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0,"",python,selection_keyboard
+65,488589,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0," ",python,content
+66,488590,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,0,"",python,selection_keyboard
+67,499039,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,15,"if __name__ == ""main"":\n",python,content
+68,500269,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"",python,selection_command
+69,502924,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"",python,selection_command
+70,503166,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"",python,selection_command
+71,503200,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"",python,selection_command
+72,503237,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_command
+73,503268,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",266,0,"",python,selection_command
+74,503296,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",267,0,"",python,selection_command
+75,503328,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",268,0,"",python,selection_command
+76,503362,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",269,0,"",python,selection_command
+77,503396,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",270,0,"",python,selection_command
+78,503431,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",271,0,"",python,selection_command
+79,503465,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",272,0,"",python,selection_command
+80,503498,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"",python,selection_command
+81,503544,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",274,0,"",python,selection_command
+82,503571,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",275,0,"",python,selection_command
+83,503605,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",276,0,"",python,selection_command
+84,503777,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",277,0,"",python,selection_command
+85,504312,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",277,0,"_",python,content
+86,504314,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",278,0,"",python,selection_keyboard
+87,504395,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",278,0,"_",python,content
+88,504397,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",279,0,"",python,selection_keyboard
+89,514849,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",278,0,"",python,selection_command
+90,515894,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",279,0,"",python,selection_command
+91,516074,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",280,0,"",python,selection_command
+92,516350,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",281,0,"",python,selection_command
+93,516729,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",282,0,"",python,selection_command
+94,517069,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",283,0,"",python,selection_command
+95,518545,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",283,0,"_",python,content
+96,518547,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",284,0,"",python,selection_keyboard
+97,527009,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,26,"if __name__ == ""__main__"":\n",python,content
+98,529298,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",288,0,"",python,selection_command
+99,552760,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,157,"",python,content
+100,557923,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,0,"    for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n        df = pd.read_parquet(parquet_file)\n        for text in df['text']:\n            print(text)\n",python,content
+101,579554,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,0,"",python,selection_command
+102,581067,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,174,"    for parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n        df = pd.read_parquet(parquet_file)\n        for text in df['text']:\n            print(text)",python,selection_command
+103,581640,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",131,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-81c2c0be-5633-4bef-b439-115d6aef8f341763318955037-2025_11_16-19.49.31.87/source.csv

@@ -0,0 +1,5 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"#!/usr/bin/env python3\n""""""\nCommon utilities for dataset serialization scripts.\n""""""\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom typing import List, Optional, Tuple, Dict\n\nimport difflib\nimport re\nimport pandas as pd\nfrom datasets import Dataset, load_dataset\n\n\n_ANSI_CSI_RE = re.compile(r""\x1b\[[0-9;?]*[ -/]*[@-~]"")\n_ANSI_OSC_TERMINATED_RE = re.compile(r""\x1b\][\s\S]*?(?:\x07|\x1b\\)"")\n_ANSI_OSC_LINE_FALLBACK_RE = re.compile(r""\x1b\][^\n]*$"")\n_BRACKETED_PASTE_ENABLE = ""\x1b[?2004h""\n_BRACKETED_PASTE_DISABLE = ""\x1b[?2004l""\n_OSC_633 = ""\x1b]633;""\n_OSC_0 = ""\x1b]0;""\n\n\n@dataclass\nclass SerializeConfig:\n    output_dir: str\n    shard_size: int\n    target_chars: int\n    overlap_chars: int\n    min_session_chars: int\n    max_docs: Optional[int]\n    long_pause_threshold_ms: int\n    csv_root: Optional[str]\n    val_ratio: float\n    arrayrecord_group_size: Optional[int] = None\n\n\ndef _clean_text(text: str) -> str:\n    # Normalize line endings and strip trailing spaces; preserve tabs/newlines.\n    return text.replace(""\r\n"", ""\n"").replace(""\r"", ""\n"").rstrip()\n\n\ndef _fenced_block(path: str, language: Optional[str], content: str) -> str:\n    lang = (language or """").lower()\n    return f""```{lang}\n{content}\n```\n""\n\n\ndef _apply_change(content: str, offset: int, length: int, new_text: str) -> str:\n    # Mirrors crowd_code_player.replay_file.apply_change\n    base = str(content)\n    text = str(new_text) if pd.notna(new_text) else """"\n    text = text.replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n    if offset > len(base):\n        base = base + ("" "" * (offset - len(base)))\n    return base[:offset] + text + base[offset + length:]\n\n\ndef _apply_backspaces(text: str) -> str:\n    out: List[str] = []\n    for ch in text:\n        if ch == ""\b"":  # \x08\n            if out:\n                out.pop()\n        else:\n            out.append(ch)\n    return """".join(out)\n\n\ndef _normalize_terminal_output(raw: str) -> str:\n    """"""\n    Normalize PTY/terminal output for training:\n      - Apply backspaces (\x08)\n      - Strip OSC (window title/shell integration) first, keeping BEL/ST terminators intact\n      - Resolve carriage returns (\r) by keeping the last rewrite per line\n      - Strip CSI (coloring etc.)\n      - Finally drop any remaining BEL (\x07)\n    """"""\n    if not raw:\n        return raw\n    s = _apply_backspaces(raw)\n    # Remove OSC sequences that are properly terminated (BEL or ST)\n    s = _ANSI_OSC_TERMINATED_RE.sub("""", s)\n    # Fallback: drop any unterminated OSC up to end-of-line only\n    s = ""\n"".join(_ANSI_OSC_LINE_FALLBACK_RE.sub("""", line) for line in s.split(""\n""))\n    # Resolve carriage returns per line:\n    # - If there are multiple rewrites, keep the last non-empty chunk\n    # - If it's CRLF (ending with '\r' before '\n'), keep the content before '\r'\n    resolved_lines: List[str] = []\n    for seg in s.split(""\n""):\n        parts = seg.split(""\r"")\n        chosen = """"\n        # pick last non-empty part if available; else last part\n        for p in reversed(parts):\n            if p != """":\n                chosen = p\n                break\n        if chosen == """" and parts:\n            chosen = parts[-1]\n        resolved_lines.append(chosen)\n    s = ""\n"".join(resolved_lines)\n    # Strip ANSI escape sequences\n    s = _ANSI_CSI_RE.sub("""", s)\n    # Remove any remaining BEL beeps\n    s = s.replace(""\x07"", """")\n    return s\n\n\ndef _line_numbered_output(content: str, start_line: Optional[int] = None, end_line: Optional[int] = None) -> str:\n    # FIXME (f.srambical): check whether this corresponds **exactly** to the output of cat -n {file_path} | sed -n '{vstart},{vend}p'\n    lines = content.splitlines()\n    total = len(lines)\n    if total == 0:\n        return """"\n    s = 1 if start_line is None else max(1, min(start_line, total))\n    e = total if end_line is None else max(1, min(end_line, total))\n    assert e >= s, ""End line number cannot be less than start line number! Likely a bug in the line numbering computation.""\n    buf: List[str] = []\n    for idx in range(s, e + 1):\n        buf.append(f""{idx:6}\t{lines[idx - 1]}"")\n    return ""\n"".join(buf)\n\n\ndef _compute_viewport(total_lines: int, center_line: int, radius: int) -> Tuple[int, int]:\n    if total_lines <= 0:\n        return (1, 0)\n    start = max(1, center_line - radius)\n    end = min(total_lines, center_line + radius)\n    assert end >= start, ""Viewport cannot have negative width! Likely a bug in the viewport computation.""\n    return (start, end)\n\n\ndef _escape_single_quotes_for_sed(text: str) -> str:\n    # Close quote, add an escaped single quote, reopen quote: '""'""'\n    return text.replace(""'"", ""'\""'\""'"")\n\n\ndef _compute_changed_block_lines(\n    before: str, after: str\n) -> Tuple[int, int, int, int, List[str]]:\n    """"""\n    Return 1-based start and end line numbers in 'before' that should be\n    replaced, 1-based start and end line numbers in 'after' that contain\n    the replacement, and the replacement lines from 'after'.\n\n    For pure deletions, the replacement list may be empty.\n    """"""\n    before_lines = before.splitlines()\n    after_lines = after.splitlines()\n    sm = difflib.SequenceMatcher(a=before_lines, b=after_lines, autojunk=False)\n    opcodes = [op for op in sm.get_opcodes() if op[0] != ""equal""]\n    assert opcodes, ""Opcode list cannot be empty! Likely a bug in the diff computation.""\n\n    first = opcodes[0]\n    last = opcodes[-1]\n    # i1/i2 refer to 'before' indices, j1/j2 to 'after'\n    start_before = max(1, first[1] + 1)\n    end_before = last[2]  # no increment since we go from 'exclusive' to 'inclusive' indexing\n    start_after = max(1, first[3] + 1)\n    end_after = last[4]\n    replacement_lines = after_lines[first[3] : last[4]]\n    return (start_before, end_before, start_after, end_after, replacement_lines)\n\n\ndef _session_to_transcript(\n    df: pd.DataFrame,\n    long_pause_threshold_ms: int,\n) -> str:\n\n    file_states: Dict[str, str] = {}\n    terminal_state: str = """"\n    per_file_event_counts: Dict[str, int] = {}\n    per_file_cursor_positions: Dict[str, Tuple[int, int]] = {}  # (offset, length) for each file\n    last_time_ms: Optional[int] = None\n\n    parts: List[str] = []\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_time: int = row[""Time""]\n        language: Optional[str] = row[""Language""]\n\n        # Long pause detection\n        if last_time_ms is not None:\n            delta = event_time - last_time_ms\n            if delta > long_pause_threshold_ms:\n                # TODO (f.srambical): think about whether we want to emit this as an observation or not\n                parts.append(f""<obs long_pause ms=\""{delta}\"" />"")\n        last_time_ms = event_time\n\n        event_type = row[""Type""]\n\n        match event_type:\n            case ""tab"":\n                # File switch event\n                parts.append(f""<act focus file=\""{file_path}\"" />"")\n                \n                # If Text is present, this is the first time opening the file\n                # and the entire file content is captured\n                text = row[""Text""]\n                if pd.notna(text):\n                    file_content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = file_content\n                    parts.append(f""// observation: file={file_path}"")\n                    parts.append(_fenced_block(file_path, language, _clean_text(file_content)))\n\n            case ""terminal_command"":\n                # Terminal command execution\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act terminal_command />"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                # Terminal output capture\n                output = row[""Text""]\n                output_str = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<obs terminal_output />"")\n                parts.append(_fenced_block(file_path, None, _clean_text(output_str)))\n\n            case ""terminal_focus"":\n                # Terminal focus event\n                parts.append(f""<act focus target=\""terminal\"" />"")\n\n            case ""git_branch_checkout"":\n                # Git branch checkout event\n                branch_info = row[""Text""]\n                branch_str = str(branch_info).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(f""<act git_branch_checkout />"")\n                parts.append(f""// git: {_clean_text(branch_str)}"")\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # Handle cursor movement\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                old_cursor = per_file_cursor_positions.get(file_path, (0, 0))\n                new_cursor = (offset, length)\n                per_file_cursor_positions[file_path] = new_cursor\n                \n                # Emit cursor movement observation if position changed\n                if old_cursor != new_cursor:\n                    parts.append(f""<act cursor file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n            case ""content"":\n                # Handle file edit events\n                offset = row[""RangeOffset""]\n                length = row[""RangeLength""]\n                new_text = row[""Text""]\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n\n                operation = ""noop""\n                if length == 0 and new_text_str:\n                    operation = ""insert""\n                elif length > 0 and not new_text_str:\n                    operation = ""delete""\n                elif length > 0 and new_text_str:\n                    operation = ""replace""\n\n                parts.append(f""<act {operation} file=\""{file_path}\"" offset=\""{offset}\"" len=\""{length}\"" />"")\n\n                if new_text_str and (operation == ""insert"" or operation == ""replace""):\n                    parts.append(_fenced_block(file_path, language, _clean_text(new_text_str)))\n\n                before = file_states.get(file_path, """")\n                after = _apply_change(before, offset, length, new_text)\n                file_states[file_path] = after\n                per_file_event_counts[file_path] = per_file_event_counts.get(file_path, 0) + 1\n\n                # Update cursor position after edit (cursor moves to end of inserted/replaced text)\n                per_file_cursor_positions[file_path] = (offset + len(new_text_str), 0)\n\n            case _:\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    return ""\n"".join(parts).strip()\n\n\ndef session_to_bash_formatted_transcript(\n    df: pd.DataFrame,\n    viewport_radius: int = 10,\n    normalize_terminal_output: bool = True,\n    coalesce_radius: int = 5,\n) -> str:\n    r""""""\n    Serialize a session to a bash-like transcript comprised of:\n      - Commands (bash fenced blocks): cat -n, sed -i 'S,Ec\...' && cat -n | sed -n 'VSTART,VENDp'\n      - Outputs (<stdout>...</stdout>) that reflect the file state after each action\n    Tracks per-file state and a per-file viewport. Viewport only shifts when selection moves out of bounds\n    or when first initialized.\n    """"""\n    file_states: Dict[str, str] = {}\n    per_file_viewport: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    parts: List[str] = []\n    terminal_output_buffer: List[str] = []\n    pending_edits_before: Dict[str, Optional[str]] = {}\n    pending_edit_regions: Dict[str, Optional[Tuple[int, int]]] = {}\n\n    def _flush_terminal_output_buffer() -> None:\n        if not terminal_output_buffer:\n            return\n        aggregated = """".join(terminal_output_buffer)\n        out = aggregated\n        if normalize_terminal_output:\n            out = _normalize_terminal_output(out)\n        cleaned = _clean_text(out)\n        if cleaned.strip():\n            parts.append(f""<stdout>\n{cleaned}\n</stdout>"")\n        terminal_output_buffer.clear()\n\n    def _flush_pending_edit_for_file(target_file: str) -> None:\n        before_snapshot = pending_edits_before.get(target_file)\n        if before_snapshot is None:\n            return\n        after_state = file_states.get(target_file, """")\n        if before_snapshot.rstrip(""\n"") == after_state.rstrip(""\n""):\n            pending_edits_before[target_file] = None\n            pending_edit_regions[target_file] = None\n            return\n        (\n            start_before,\n            end_before,\n            start_after,\n            end_after,\n            repl_lines,\n        ) = _compute_changed_block_lines(before_snapshot, after_state)\n        before_total_lines = len(before_snapshot.splitlines())\n        if end_before < start_before:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            if start_before <= max(1, before_total_lines):\n                sed_cmd = f""sed -i '{start_before}i\\\n{sed_payload}' {target_file}""\n            else:\n                sed_cmd = f""sed -i '$a\\\n{sed_payload}' {target_file}""\n        elif not repl_lines:\n            sed_cmd = f""sed -i '{start_before},{end_before}d' {target_file}""\n        else:\n            escaped_lines = [_escape_single_quotes_for_sed(line) for line in repl_lines]\n            sed_payload = ""\n"".join(escaped_lines)\n            sed_cmd = f""sed -i '{start_before},{end_before}c\\\n{sed_payload}' {target_file}""\n        total_lines = len(after_state.splitlines())\n        center = (start_after + end_after) // 2\n        vp = _compute_viewport(total_lines, center, viewport_radius)\n        per_file_viewport[target_file] = vp\n        vstart, vend = vp\n        chained_cmd = f""{sed_cmd} && cat -n {target_file} | sed -n '{vstart},{vend}p'""\n        parts.append(_fenced_block(target_file, ""bash"", _clean_text(chained_cmd)))\n        viewport_output = _line_numbered_output(after_state, vstart, vend)\n        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n        pending_edits_before[target_file] = None\n        pending_edit_regions[target_file] = None\n\n    def _flush_all_pending_edits() -> None:\n        for fname in list(pending_edits_before.keys()):\n            _flush_pending_edit_for_file(fname)\n\n    for i in range(len(df)):\n        row = df.iloc[i]\n        file_path: str = row[""File""]\n        event_type = row[""Type""]\n        \n        match event_type:\n            case ""tab"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                text = row[""Text""]\n                if pd.notna(text):\n                    content = str(text).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                    file_states[file_path] = content\n                    # First open with full file capture\n                    cmd = f""cat -n {file_path}""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    output = _line_numbered_output(content)\n                    parts.append(f""<stdout>\n{output}\n</stdout>"")\n                else:\n                    # File switch without content snapshot: show current viewport only\n                    content = file_states.get(file_path, """")\n                    total_lines = len(content.splitlines())\n                    vp = per_file_viewport.get(file_path)\n                    if not vp or vp[1] == 0:\n                        vp = _compute_viewport(total_lines, 1, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                    if vp:\n                        vstart, vend = vp\n                        cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                        parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                        viewport_output = _line_numbered_output(content, vstart, vend)\n                        parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""content"":\n                _flush_terminal_output_buffer()\n                offset = int(row[""RangeOffset""])\n                length = int(row[""RangeLength""])\n                new_text = row[""Text""]\n                before = file_states.get(file_path, """")\n                # Approximate current edit region in line space\n                new_text_str = str(new_text) if pd.notna(new_text) else """"\n                start_line_current = before[:offset].count(""\n"") + 1\n                deleted_chunk = before[offset:offset + length]\n                lines_added = new_text_str.count(""\n"")\n                lines_deleted = deleted_chunk.count(""\n"")\n                region_start = start_line_current\n                region_end = start_line_current + max(lines_added, lines_deleted, 0)\n                # Flush pending edits if this edit is far from the pending region\n                current_region = pending_edit_regions.get(file_path)\n                if current_region is not None:\n                    rstart, rend = current_region\n                    if region_start < (rstart - coalesce_radius) or region_start > (rend + coalesce_radius):\n                        _flush_pending_edit_for_file(file_path)\n                        current_region = None\n                after = _apply_change(before, offset, length, new_text)\n                if pending_edits_before.get(file_path) is None:\n                    pending_edits_before[file_path] = before\n                # Update/initialize region union\n                if current_region is None:\n                    pending_edit_regions[file_path] = (region_start, max(region_start, region_end))\n                else:\n                    rstart, rend = current_region\n                    pending_edit_regions[file_path] = (min(rstart, region_start), max(rend, region_end))\n                file_states[file_path] = after\n\n            case ""selection_command"" | ""selection_mouse"" | ""selection_keyboard"":\n                # During an edit burst (pending edits), suppress flush and viewport emissions\n                if pending_edits_before.get(file_path) is None:\n                    _flush_terminal_output_buffer()\n                else:\n                    # Skip emitting viewport while edits are pending to avoid per-keystroke sed/cat spam\n                    continue\n                offset = int(row[""RangeOffset""])\n                content = file_states.get(file_path, """")\n                total_lines = len(content.splitlines())\n                target_line = content[:offset].count(""\n"") + 1\n                vp = per_file_viewport.get(file_path)\n                should_emit = False\n                if not vp or vp[1] == 0:\n                    vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                    per_file_viewport[file_path] = vp\n                    should_emit = True\n                else:\n                    vstart, vend = vp\n                    if target_line < vstart or target_line > vend:\n                        vp = _compute_viewport(total_lines, target_line, viewport_radius)\n                        per_file_viewport[file_path] = vp\n                        should_emit = True\n                if should_emit and vp:\n                    vstart, vend = vp\n                    cmd = f""cat -n {file_path} | sed -n '{vstart},{vend}p'""\n                    parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n                    viewport_output = _line_numbered_output(content, vstart, vend)\n                    parts.append(f""<stdout>\n{viewport_output}\n</stdout>"")\n\n            case ""terminal_command"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                command = row[""Text""]\n                command_str = str(command).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(command_str)))\n\n            case ""terminal_output"":\n                output = row[""Text""]\n                raw_output = str(output).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                terminal_output_buffer.append(raw_output)\n\n            case ""terminal_focus"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                # No-op for bash transcript; focus changes don't emit commands/output\n                pass\n\n            case ""git_branch_checkout"":\n                _flush_all_pending_edits()\n                _flush_terminal_output_buffer()\n                branch_info = row[""Text""]\n                branch_str = str(branch_info).replace(""\\n"", ""\n"").replace(""\\r"", ""\r"")\n                cleaned = _clean_text(branch_str)\n                m = re.search(r""to '([^']+)'"", cleaned)\n                if not m:\n                    raise ValueError(f""Could not extract branch name from git checkout message: {cleaned}"")\n                branch_name = m.group(1).strip()\n                # Safe-quote branch if it contains special characters\n                if re.search(r""[^A-Za-z0-9._/\\-]"", branch_name):\n                    branch_name = ""'"" + branch_name.replace(""'"", ""'\""'\""'"") + ""'""\n                cmd = f""git checkout {branch_name}""\n                parts.append(_fenced_block(file_path, ""bash"", _clean_text(cmd)))\n\n            case _:\n                raise ValueError(f""Unknown event type: {event_type}"")\n\n    _flush_all_pending_edits()\n    _flush_terminal_output_buffer()\n    return ""\n"".join(parts).strip()\n\ndef load_hf_csv(hf_path: str, split: str) -> Dataset:\n    loaded = load_dataset(hf_path, split=split)\n\n    assert isinstance(loaded, Dataset), ""Expected a Dataset from load_dataset""\n    return loaded\n\n\ndef _discover_local_sessions(root: Path) -> List[Path]:\n    # Recursively find all CSV files\n    paths: List[Path] = []\n    for p in root.rglob(""*.csv""):\n        if p.is_file():\n            paths.append(p)\n    paths.sort()\n    return paths\n\n\ndef _chunk_text(text: str, target_chars: int, overlap_chars: int) -> List[str]:\n    """"""Split a long text into overlapping chunks near target length.""""""\n    if target_chars <= 0:\n        return [text]\n    n = len(text)\n    if n <= target_chars:\n        return [text]\n\n    chunks: List[str] = []\n    start = 0\n    # Ensure sane overlap\n    overlap = max(0, min(overlap_chars, target_chars // 2))\n    while start < n:\n        end_target = min(start + target_chars, n)\n        if end_target < n:\n            end = end_target\n        else:\n            end = n\n        chunk = text[start:end].strip()\n        chunks.append(chunk)\n        if end == n:\n            break\n        # advance with overlap\n        start = max(0, end - overlap)\n        if start >= n:\n            break\n    return chunks\n\n\n",python,tab
+2,1748,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"7:49:31 PM [info] Activating crowd-code\n7:49:31 PM [info] Recording started\n7:49:31 PM [info] Initializing git provider using file system watchers...\n7:49:32 PM [info] Git repository found\n",Log,tab
+3,1981,"extension-output-pdoom-org.crowd-code-#1-crowd-code",189,0,"7:49:32 PM [info] Git provider initialized successfully\n7:49:32 PM [info] Initial git state: [object Object]\n",Log,content
+4,163843,"crowd-pilot/crowd-pilot/serialization_utils.py",0,0,"",python,tab

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-8c6fe2f6-5c87-474b-8190-b29853bcc7151755414648501-2025_08_17-09.10.53.985/source.csv

@@ -0,0 +1,4 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,2,"big_vision/train.py",0,0,"# Copyright 2024 Big Vision Authors.\n#\n# Licensed under the Apache License, Version 2.0 (the ""License"");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n#     http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an ""AS IS"" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\n""""""Training loop example.\n\nThis is a basic variant of a training loop, good starting point for fancy ones.\n""""""\n# pylint: disable=consider-using-from-import\n# pylint: disable=logging-fstring-interpolation\n\nimport functools\nimport importlib\nimport multiprocessing.pool\nimport os\n\nfrom absl import app\nfrom absl import flags\nfrom absl import logging\nimport big_vision.evaluators.common as eval_common\nimport big_vision.input_pipeline as input_pipeline\nimport big_vision.optax as bv_optax\nimport big_vision.sharding as bv_sharding\nimport big_vision.utils as u\nfrom clu import parameter_overview\nimport flax.linen as nn\nimport jax\nfrom jax.experimental import multihost_utils\nfrom jax.experimental.array_serialization import serialization as array_serial\nfrom jax.experimental.shard_map import shard_map\nimport jax.numpy as jnp\nfrom ml_collections import config_flags\nimport numpy as np\nimport optax\nimport tensorflow as tf\n\nfrom tensorflow.io import gfile\n\n\nconfig_flags.DEFINE_config_file(\n    ""config"", None, ""Training configuration."", lock_config=True)\n\nflags.DEFINE_string(""workdir"", default=None, help=""Work unit directory."")\nflags.DEFINE_boolean(""cleanup"", default=False,\n                     help=""Delete workdir (only) after successful completion."")\n\n# Adds jax flags to the program.\njax.config.parse_flags_with_absl()\n# Transfer guard will fail the program whenever that data between a host and\n# a device is transferred implicitly. This often catches subtle bugs that\n# cause slowdowns and memory fragmentation. Explicit transfers are done\n# with jax.device_put and jax.device_get.\njax.config.update(""jax_transfer_guard"", ""disallow"")\n# Fixes design flaw in jax.random that may cause unnecessary d2d comms.\njax.config.update(""jax_threefry_partitionable"", True)\n\n\nNamedSharding = jax.sharding.NamedSharding\nP = jax.sharding.PartitionSpec\n\n\ndef main(argv):\n  del argv\n\n  # This is needed on multihost systems, but crashes on non-TPU single-host.\n  if os.environ.get(""BV_JAX_INIT""):\n    jax.distributed.initialize()\n\n  # Make sure TF does not touch GPUs.\n  tf.config.set_visible_devices([], ""GPU"")\n\n  config = flags.FLAGS.config\n\n################################################################################\n#                                                                              #\n#                                Set up logging                                #\n#                                                                              #\n################################################################################\n\n  # Set up work directory and print welcome message.\n  workdir = flags.FLAGS.workdir\n  logging.info(\n      f""\u001b[33mHello from process {jax.process_index()} holding ""\n      f""{jax.local_device_count()}/{jax.device_count()} devices and ""\n      f""writing to workdir {workdir}.\u001b[0m"")\n  logging.info(f""The config:\n{config}"")\n\n  save_ckpt_path = None\n  if workdir:  # Always create if requested, even if we may not write into it.\n    gfile.makedirs(workdir)\n    save_ckpt_path = os.path.join(workdir, ""checkpoint.bv"")\n\n  # The pool is used to perform misc operations such as logging in async way.\n  pool = multiprocessing.pool.ThreadPool(1)\n\n  # Here we register preprocessing ops from modules listed on `pp_modules`.\n  for m in config.get(""pp_modules"", [""ops_general"", ""ops_image"", ""ops_text""]):\n    importlib.import_module(f""big_vision.pp.{m}"")\n\n  # Setup up logging and experiment manager.\n  xid, wid = -1, -1\n  fillin = lambda s: s\n  def info(s, *a):\n    logging.info(""\u001b[33mNOTE\u001b[0m: "" + s, *a)\n  def write_note(note):\n    if jax.process_index() == 0:\n      info(""%s"", note)\n\n  mw = u.BigVisionMetricWriter(xid, wid, workdir, config)\n\n  # Allow for things like timings as early as possible!\n  u.chrono.inform(measure=mw.measure, write_note=write_note)\n\n################################################################################\n#                                                                              #\n#                                Set up Mesh                                   #\n#                                                                              #\n################################################################################\n\n  # We rely on jax mesh_utils to organize devices, such that communication\n  # speed is the fastest for the last dimension, second fastest for the\n  # penultimate dimension, etc.\n  config_mesh = config.get(""mesh"", [(""data"", jax.device_count())])\n\n  # Sharding rules with default\n  sharding_rules = config.get(""sharding_rules"", [(""act_batch"", ""data"")])\n\n  write_note(""Creating device mesh..."")\n  mesh = u.create_device_mesh(\n      config_mesh,\n      allow_split_physical_axes=config.get(""mesh_allow_split_physical_axes"",\n                                           False))\n  repl_sharding = jax.sharding.NamedSharding(mesh, P())\n\n  # Consistent device order is important to ensure correctness of various train\n  # loop components, such as input pipeline, update step, evaluators. The\n  # order presribed by the `devices_flat` variable should be used throughout\n  # the program.\n  devices_flat = mesh.devices.flatten()\n\n################################################################################\n#                                                                              #\n#                                Input Pipeline                                #\n#                                                                              #\n################################################################################\n\n  write_note(""Initializing train dataset..."")\n  batch_size = config.input.batch_size\n  if batch_size % jax.device_count() != 0:\n    raise ValueError(f""Batch size ({batch_size}) must ""\n                     f""be divisible by device number ({jax.device_count()})"")\n  info(""Global batch size %d on %d hosts results in %d local batch size. With ""\n       ""%d dev per host (%d dev total), that's a %d per-device batch size."",\n       batch_size, jax.process_count(), batch_size // jax.process_count(),\n       jax.local_device_count(), jax.device_count(),\n       batch_size // jax.device_count())\n\n  train_ds, ntrain_img = input_pipeline.training(config.input)\n\n  total_steps = u.steps(""total"", config, ntrain_img, batch_size)\n  def get_steps(name, default=ValueError, cfg=config):\n    return u.steps(name, cfg, ntrain_img, batch_size, total_steps, default)\n\n  u.chrono.inform(total_steps=total_steps, global_bs=batch_size,\n                  steps_per_epoch=ntrain_img / batch_size)\n\n  info(""Running for %d steps, that means %f epochs"",\n       total_steps, total_steps * batch_size / ntrain_img)\n\n  # Start input pipeline as early as possible.\n  n_prefetch = config.get(""prefetch_to_device"", 1)\n  train_iter = input_pipeline.start_global(train_ds, devices_flat, n_prefetch)\n\n################################################################################\n#                                                                              #\n#                           Create Model & Optimizer                           #\n#                                                                              #\n################################################################################\n\n  write_note(""Creating model..."")\n  model_mod = importlib.import_module(f""big_vision.models.{config.model_name}"")\n  model = model_mod.Model(\n      num_classes=config.num_classes, **config.get(""model"", {}))\n\n  def init(rng):\n    batch = jax.tree.map(lambda x: jnp.zeros(x.shape, x.dtype.as_numpy_dtype),\n                         train_ds.element_spec)\n    params = model.init(rng, batch[""image""])[""params""]\n\n    # Set bias in the head to a low value, such that loss is small initially.\n    if ""init_head_bias"" in config:\n      params[""head""][""bias""] = jnp.full_like(params[""head""][""bias""],\n                                             config[""init_head_bias""])\n\n    return params\n\n  # This seed makes the Jax part of things (like model init) deterministic.\n  # However, full training still won't be deterministic, for example due to the\n  # tf.data pipeline not being deterministic even if we would set TF seed.\n  # See (internal link) for a fun read on what it takes.\n  rng = jax.random.PRNGKey(u.put_cpu(config.get(""seed"", 0)))\n\n  write_note(""Inferring parameter shapes..."")\n  rng, rng_init = jax.random.split(rng)\n  params_shape = jax.eval_shape(init, rng_init)\n\n  write_note(""Inferring optimizer state shapes..."")\n  tx, sched_fns = bv_optax.make(config, nn.unbox(params_shape), sched_kw=dict(\n      total_steps=total_steps, batch_size=batch_size, data_size=ntrain_img))\n  opt_shape = jax.eval_shape(tx.init, params_shape)\n  # We jit this, such that the arrays are created on the CPU, not device[0].\n  sched_fns_cpu = [u.jit_cpu()(sched_fn) for sched_fn in sched_fns]\n\n  if jax.process_index() == 0:\n    num_params = sum(np.prod(p.shape) for p in jax.tree.leaves(params_shape))\n    mw.measure(""num_params"", num_params)\n\n################################################################################\n#                                                                              #\n#                               Shard & Transfer                               #\n#                                                                              #\n################################################################################\n\n  write_note(""Inferring shardings..."")\n  train_state_shape = {""params"": params_shape, ""opt"": opt_shape}\n\n  strategy = config.get(""sharding_strategy"", [("".*"", ""replicate"")])\n  with nn.logical_axis_rules(sharding_rules):\n    train_state_sharding = bv_sharding.infer_sharding(\n        train_state_shape, strategy=strategy, mesh=mesh)\n\n  write_note(""Transferring train_state to devices..."")\n  # RNG is always replicated\n  rng_init = u.reshard(rng_init, repl_sharding)\n\n  # Parameters and the optimizer are now global (distributed) jax arrays.\n  params = jax.jit(init, out_shardings=train_state_sharding[""params""])(rng_init)\n  opt = jax.jit(tx.init, out_shardings=train_state_sharding[""opt""])(params)\n\n  rng, rng_loop = jax.random.split(rng, 2)\n  rng_loop = u.reshard(rng_loop, repl_sharding)\n  del rng  # not used anymore, so delete it.\n\n  # At this point we have everything we need to form a train state. It contains\n  # all the parameters that are passed and updated by the main training step.\n  # From here on, we have no need for Flax AxisMetadata (such as partitioning).\n  train_state = nn.unbox({""params"": params, ""opt"": opt})\n  del params, opt  # Delete to avoid memory leak or accidental reuse.\n\n  write_note(""Logging parameter overview..."")\n  parameter_overview.log_parameter_overview(\n      train_state[""params""], msg=""Init params"",\n      include_stats=""global"", jax_logging_process=0)\n\n################################################################################\n#                                                                              #\n#                                 Update Step                                  #\n#                                                                              #\n################################################################################\n\n  @functools.partial(\n      jax.jit,\n      donate_argnums=(0,),\n      out_shardings=(train_state_sharding, repl_sharding))\n  def update_fn(train_state, rng, batch):\n    """"""Update step.""""""\n\n    images, labels = batch[""image""], batch[""labels""]\n\n    step_count = bv_optax.get_count(train_state[""opt""], jittable=True)\n    rng = jax.random.fold_in(rng, step_count)\n\n    if config.get(""mixup"") and config.mixup.p:\n      # The shard_map below makes mixup run on every device independently and\n      # thus avoids unnecessary communication.\n      sharded_mixup_fn = shard_map(\n          u.get_mixup(rng, config.mixup.p),\n          mesh=jax.sharding.Mesh(devices_flat, (""data"",)),\n          in_specs=P(""data""), out_specs=(P(), P(""data""), P(""data"")))\n      rng, (images, labels), _ = sharded_mixup_fn(images, labels)\n\n    # Get device-specific loss rng.\n    rng, rng_model = jax.random.split(rng, 2)\n\n    def loss_fn(params):\n      logits, _ = model.apply(\n          {""params"": params}, images,\n          train=True, rngs={""dropout"": rng_model})\n      return getattr(u, config.get(""loss"", ""sigmoid_xent""))(\n          logits=logits, labels=labels)\n\n    params, opt = train_state[""params""], train_state[""opt""]\n    loss, grads = jax.value_and_grad(loss_fn)(params)\n    updates, opt = tx.update(grads, opt, params)\n    params = optax.apply_updates(params, updates)\n\n    measurements = {""training_loss"": loss}\n    gs = jax.tree.leaves(bv_optax.replace_frozen(config.schedule, grads, 0.))\n    measurements[""l2_grads""] = jnp.sqrt(sum([jnp.sum(g * g) for g in gs]))\n    ps = jax.tree.leaves(params)\n    measurements[""l2_params""] = jnp.sqrt(sum([jnp.sum(p * p) for p in ps]))\n    us = jax.tree.leaves(updates)\n    measurements[""l2_updates""] = jnp.sqrt(sum([jnp.sum(u * u) for u in us]))\n\n    return {""params"": params, ""opt"": opt}, measurements\n\n################################################################################\n#                                                                              #\n#                               Load Checkpoint                                #\n#                                                                              #\n################################################################################\n\n  # Decide how to initialize training. The order is important.\n  # 1. Always resumes from the existing checkpoint, e.g. resumes a finetune job.\n  # 2. Resume from a previous checkpoint, e.g. start a cooldown training job.\n  # 3. Initialize model from something, e,g, start a fine-tuning job.\n  # 4. Train from scratch.\n  resume_ckpt_path = None\n  if save_ckpt_path and gfile.exists(f""{save_ckpt_path}-LAST""):\n    resume_ckpt_path = save_ckpt_path\n  elif config.get(""resume""):\n    resume_ckpt_path = fillin(config.resume)\n\n  ckpt_mngr = None\n  if save_ckpt_path or resume_ckpt_path:\n    ckpt_mngr = array_serial.GlobalAsyncCheckpointManager()\n\n  if resume_ckpt_path:\n    write_note(f""Resuming training from checkpoint {resume_ckpt_path}..."")\n    jax.tree.map(lambda x: x.delete(), train_state)\n    del train_state\n    shardings = {\n        **train_state_sharding,\n        ""chrono"": jax.tree.map(lambda _: repl_sharding,\n                               u.chrono.save()),\n    }\n    loaded = u.load_checkpoint_ts(\n        resume_ckpt_path, tree=shardings, shardings=shardings)\n    train_state = {key: loaded[key] for key in train_state_sharding.keys()}\n\n    u.chrono.load(jax.device_get(loaded[""chrono""]))\n    del loaded\n  elif config.get(""model_init""):\n    write_note(f""Initialize model from {config.model_init}..."")\n    # TODO: when updating the `load` API soon, do pass and request the\n    # full `train_state` from it. Examples where useful: VQVAE, BN.\n    train_state[""params""] = model_mod.load(\n        train_state[""params""], config.model_init, config.get(""model""),\n        **config.get(""model_load"", {}))\n\n    # load has the freedom to return params not correctly sharded. Think of for\n    # example ViT resampling position embedings on CPU as numpy arrays.\n    train_state[""params""] = u.reshard(\n        train_state[""params""], train_state_sharding[""params""])\n\n    parameter_overview.log_parameter_overview(\n        train_state[""params""], msg=""restored params"",\n        include_stats=""global"", jax_logging_process=0)\n\n\n################################################################################\n#                                                                              #\n#                                 Setup Evals                                  #\n#                                                                              #\n################################################################################\n\n  # We do not jit/pmap this function, because it is passed to evaluator that\n  # does it later. We output as many intermediate tensors as possible for\n  # maximal flexibility. Later `jit` will prune out things that are not needed.\n  def eval_logits_fn(train_state, batch):\n    logits, out = model.apply({""params"": train_state[""params""]}, batch[""image""])\n    return logits, out\n\n  def eval_loss_fn(train_state, batch):\n    logits, _ = model.apply({""params"": train_state[""params""]}, batch[""image""])\n    loss_fn = getattr(u, config.get(""loss"", ""sigmoid_xent""))\n    return {\n        ""loss"": loss_fn(logits=logits, labels=batch[""labels""], reduction=False)\n    }\n\n  eval_fns = {\n      ""predict"": eval_logits_fn,\n      ""loss"": eval_loss_fn,\n  }\n\n  # Only initialize evaluators when they are first needed.\n  @functools.lru_cache(maxsize=None)\n  def evaluators():\n    return eval_common.from_config(\n        config, eval_fns,\n        lambda s: write_note(f""Init evaluator: {s}…\n{u.chrono.note}""),\n        lambda key, cfg: get_steps(key, default=None, cfg=cfg),\n        devices_flat,\n    )\n\n  # At this point we need to know the current step to see whether to run evals.\n  write_note(""Inferring the first step number..."")\n  first_step_device = bv_optax.get_count(train_state[""opt""], jittable=True)\n  first_step = int(jax.device_get(first_step_device))\n  u.chrono.inform(first_step=first_step)\n\n  # Note that training can be pre-empted during the final evaluation (i.e.\n  # just after the final checkpoint has been written to disc), in which case we\n  # want to run the evals.\n  if first_step in (total_steps, 0):\n    write_note(""Running initial or final evals..."")\n    mw.step_start(first_step)\n    for (name, evaluator, _, prefix) in evaluators():\n      if config.evals[name].get(""skip_first"") and first_step != total_steps:\n        continue\n      write_note(f""{name} evaluation...\n{u.chrono.note}"")\n      with u.chrono.log_timing(f""z/secs/eval/{name}""):\n        with mesh, nn.logical_axis_rules(sharding_rules):\n          for key, value in evaluator.run(train_state):\n            mw.measure(f""{prefix}{key}"", value)\n\n################################################################################\n#                                                                              #\n#                                  Train Loop                                  #\n#                                                                              #\n################################################################################\n\n  prof = None  # Keeps track of start/stop of profiler state.\n\n  write_note(""Starting training loop, compiling the first step..."")\n  for step, batch in zip(range(first_step + 1, total_steps + 1), train_iter):\n    mw.step_start(step)\n\n    with jax.profiler.StepTraceAnnotation(""train_step"", step_num=step):\n      with u.chrono.log_timing(""z/secs/update0"", noop=step > first_step + 1):\n        with mesh, nn.logical_axis_rules(sharding_rules):\n          train_state, measurements = update_fn(train_state, rng_loop, batch)\n\n    # On the first host, let's always profile a handful of early steps.\n    if jax.process_index() == 0:\n      prof = u.startstop_prof(prof, step, first_step, get_steps(""log_training""))\n\n    # Report training progress\n    if (u.itstime(step, get_steps(""log_training""), total_steps, host=0)\n        or u.chrono.warmup and jax.process_index() == 0):\n      for i, sched_fn_cpu in enumerate(sched_fns_cpu):\n        mw.measure(f""global_schedule{i if i else ''}"",\n                   sched_fn_cpu(u.put_cpu(step - 1)))\n      measurements = jax.device_get(measurements)\n      for name, value in measurements.items():\n        mw.measure(name, value)\n      u.chrono.tick(step)\n      for k in (""training_loss"", ""l2_grads"", ""l2_updates"", ""l2_params""):\n        if not np.isfinite(measurements.get(k, 0.0)):\n          raise RuntimeError(f""{k} became nan or inf somewhere within steps ""\n                             f""[{step - get_steps('log_training')}, {step}]"")\n\n    # Checkpoint saving\n    keep_last = total_steps if get_steps(""ckpt"", None) else None\n    keep_ckpt_steps = get_steps(""keep_ckpt"", None) or keep_last\n    if save_ckpt_path and (\n        (keep := u.itstime(step, keep_ckpt_steps, total_steps, first=False))\n        or u.itstime(step, get_steps(""ckpt"", None), total_steps, first=True)\n    ):\n      u.chrono.pause(wait_for=train_state)\n\n      # Copy because we add extra stuff to the checkpoint.\n      ckpt = {**train_state}\n\n      # To save chrono state correctly and safely in a multihost setup, we\n      # broadcast the state to all hosts and convert it to a global array.\n      with jax.transfer_guard(""allow""):\n        chrono_ckpt = multihost_utils.broadcast_one_to_all(u.chrono.save())\n      chrono_shardings = jax.tree.map(lambda _: repl_sharding, chrono_ckpt)\n      ckpt = ckpt | {""chrono"": u.reshard(chrono_ckpt, chrono_shardings)}\n\n      u.save_checkpoint_ts(ckpt_mngr, ckpt, save_ckpt_path, step, keep)\n      u.chrono.resume()\n\n    for (name, evaluator, log_steps, prefix) in evaluators():\n      if u.itstime(step, log_steps, total_steps, first=False, last=True):\n        u.chrono.pause(wait_for=train_state)\n        u.chrono.tick(step)  # Record things like epoch number, core hours etc.\n        write_note(f""{name} evaluation...\n{u.chrono.note}"")\n        with u.chrono.log_timing(f""z/secs/eval/{name}""):\n          with mesh, nn.logical_axis_rules(sharding_rules):\n            for key, value in evaluator.run(train_state):\n              mw.measure(f""{prefix}{key}"", jax.device_get(value))\n        u.chrono.resume()\n    mw.step_end()\n\n  # Always give a chance to stop the profiler, no matter how things ended.\n  # TODO: can we also do this when dying of an exception like OOM?\n  if jax.process_index() == 0 and prof is not None:\n    u.startstop_prof(prof)\n\n  # Last note needs to happen before the pool's closed =)\n  write_note(f""Done!\n{u.chrono.note}"")\n\n  pool.close()\n  pool.join()\n  mw.close()\n  if ckpt_mngr:\n    ckpt_mngr.wait_until_finished()\n\n  # Make sure all hosts stay up until the end of main.\n  u.sync()\n\n  u.maybe_cleanup_workdir(workdir, flags.FLAGS.cleanup, info)\n\n\nif __name__ == ""__main__"":\n  app.run(main)\n",python,tab
+2,96,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"9:10:53 AM [info] Activating crowd-code\n9:10:53 AM [info] Recording started\n9:10:53 AM [info] Initializing git provider using file system watchers...\n9:10:54 AM [info] Git repository found\n9:10:54 AM [info] Git provider initialized successfully\n",Log,tab
+3,196,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"9:10:54 AM [info] Initial git state: [object Object]\n",Log,content

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-e5295f7c-e65d-4455-aaa4-44e96e5467f81765288393722-2025_12_09-14.53.21.307/source.csv

@@ -0,0 +1,57 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,537,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"2:53:21 PM [info] Activating crowd-code\n2:53:21 PM [info] Recording started\n2:53:21 PM [info] Initializing git provider using file system watchers...\n2:53:21 PM [info] No workspace folder found\n",Log,tab
+3,2797,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",0,0,"#!/usr/bin/env python3\nimport sys\nfrom pathlib import Path\nimport pandas as pd\n\ninput_dir = sys.argv[1]\nfor parquet_file in sorted(Path(input_dir).glob(""shard_*.parquet"")):\n    df = pd.read_parquet(parquet_file)\n    for text in df['text']:\n        print(text)\n",python,tab
+4,6424,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",240,0,"",python,selection_command
+5,6630,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+6,6886,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",240,0,"",python,selection_command
+7,7089,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+8,7515,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"\n",python,content
+9,10888,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"i",python,content
+10,10891,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"",python,selection_keyboard
+11,11118,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",262,0,"f",python,content
+12,11120,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0,"",python,selection_keyboard
+13,11129,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",263,0," ",python,content
+14,11133,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"",python,selection_keyboard
+15,14382,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",264,0,"_",python,content
+16,14385,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",265,0,"",python,selection_keyboard
+17,16160,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,4,"if __name__ == ""__main__"":\n",python,content
+18,18110,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",288,0,"    main()\n",python,content
+19,30407,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",288,0,"",python,selection_command
+20,30562,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"",python,selection_command
+21,30809,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",260,0,"",python,selection_command
+22,30841,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",240,0,"",python,selection_command
+23,30877,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",212,0,"",python,selection_command
+24,30910,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",173,0,"",python,selection_command
+25,30943,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",104,0,"",python,selection_command
+26,30971,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,0,"",python,selection_command
+27,31005,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",79,0,"",python,selection_command
+28,31145,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",59,0,"",python,selection_command
+29,31383,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",79,0,"",python,selection_command
+30,35054,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",79,0,"\n",python,content
+31,38264,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,0,"d",python,content
+32,38266,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",81,0,"",python,selection_keyboard
+33,40019,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,2,"def main():\n",python,content
+34,41614,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,0,"",python,selection_command
+35,42532,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",91,0,"\n    ",python,content
+36,45322,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",92,4,"",python,content
+37,226426,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",93,0,"",python,selection_command
+38,227609,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",92,0,"",python,selection_command
+39,227801,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",80,0,"",python,selection_command
+40,231987,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",92,0,"",python,selection_command
+41,233629,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",93,0,"",python,selection_command
+42,233634,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",117,0,"",python,selection_command
+43,235362,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",186,0,"",python,selection_command
+44,236390,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",190,0,"",python,selection_command
+45,237568,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",229,0,"",python,selection_command
+46,238498,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",257,0,"",python,selection_command
+47,238609,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"",python,selection_command
+48,241015,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"",python,selection_command
+49,241191,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",282,0,"",python,selection_command
+50,241313,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",273,0,"",python,selection_command
+51,241541,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",261,0,"",python,selection_command
+52,241670,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",233,0,"",python,selection_command
+53,241849,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",194,0,"",python,selection_command
+54,275561,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",125,0,"",python,selection_command
+55,275724,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",101,0,"",python,selection_command
+56,275887,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",92,0,"",python,selection_command
+57,276025,"/home/franz.srambical/crowd-pilot/crowd_pilot/read_dataset.py",88,0,"",python,selection_command

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-e6fd64e1-70f6-461f-bc3d-ac75f6654a311756111953368-2025_08_25-10.52.42.401/source.csv

@@ -0,0 +1,119 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,3,"/fast/home/franz.srambical/jafar/train_tokenizer.py",0,0,"import os\n\n# os.environ['XLA_FLAGS'] = (\n#     '--xla_python_client_mem_fraction=.98 '\n# )\n# FIXME (f.srambical): test whether this increases throughput\nos.environ['XLA_FLAGS'] = (\n    '--xla_gpu_enable_latency_hiding_scheduler=true '\n    '--xla_gpu_enable_async_collectives=true '\n)\n\nfrom dataclasses import dataclass, field\nfrom typing import cast, Optional\n\nimport einops\nimport itertools\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    init_lr: float = 0.0\n    max_lr: float = 3e-4\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        20000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[TokenizerVQVAE, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        TokenizerVQVAE(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(\n    model: TokenizerVQVAE, args: Args\n) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    grain_iterator: grain.DataLoaderIterator,\n    restore_step: Optional[int] = None,\n) -> tuple[int, nnx.Optimizer, grain.DataLoaderIterator]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            restore_step,\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, grain_iterator\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    tokenizer, rng = build_model(args, rng)\n\n    _, params, _ = nnx.split(tokenizer, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(tokenizer, args)\n    del tokenizer\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    mesh, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    grain_iterator = build_dataloader(args)\n\n    # --- Restore checkpoint ---\n    step, optimizer, grain_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, grain_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def tokenizer_loss_fn(\n        model: TokenizerVQVAE, inputs: dict\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        model.train()\n        outputs = model(inputs, training=True)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        mse = jnp.square(gt - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_clipped = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_clipped, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_clipped, recon)).mean()\n        _, index_counts = jnp.unique_counts(\n            jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n        )\n        codebook_usage = (index_counts != 0).mean()\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=codebook_usage,\n        )\n        return loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: TokenizerVQVAE) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            return tokenizer_loss_fn(model, inputs)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""encoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""encoder""]\n            )\n            metrics[""vq_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""vq""]\n            )\n            metrics[""decoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""decoder""]\n            )\n        return loss, recon, metrics\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    if jax.process_index() == 0:\n        first_videos = next(dataloader)\n        sample_inputs = dict(videos=first_videos)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader = itertools.chain([first_videos], dataloader)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            inputs = dict(videos=videos)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+2,302,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:52:42 AM [info] Activating crowd-code\n10:52:42 AM [info] Recording started\n10:52:42 AM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,811,"extension-output-pdoom-org.crowd-code-#1-crowd-code",153,0,"10:52:42 AM [info] Git repository found\n10:52:42 AM [info] Git provider initialized successfully\n10:52:42 AM [info] Initial git state: [object Object]\n",Log,content
+4,4248,"/fast/home/franz.srambical/jafar/train_tokenizer.py",0,0,"",python,tab
+5,8874,"/fast/home/franz.srambical/jafar/train_tokenizer.py",392,0,"",python,selection_command
+6,9119,"/fast/home/franz.srambical/jafar/train_tokenizer.py",375,0,"",python,selection_command
+7,9169,"/fast/home/franz.srambical/jafar/train_tokenizer.py",361,0,"",python,selection_command
+8,9186,"/fast/home/franz.srambical/jafar/train_tokenizer.py",360,0,"",python,selection_command
+9,9234,"/fast/home/franz.srambical/jafar/train_tokenizer.py",326,0,"",python,selection_command
+10,9302,"/fast/home/franz.srambical/jafar/train_tokenizer.py",285,0,"",python,selection_command
+11,9302,"/fast/home/franz.srambical/jafar/train_tokenizer.py",284,0,"",python,selection_command
+12,9314,"/fast/home/franz.srambical/jafar/train_tokenizer.py",282,0,"",python,selection_command
+13,9366,"/fast/home/franz.srambical/jafar/train_tokenizer.py",235,0,"",python,selection_command
+14,9401,"/fast/home/franz.srambical/jafar/train_tokenizer.py",181,0,"",python,selection_command
+15,9415,"/fast/home/franz.srambical/jafar/train_tokenizer.py",153,0,"",python,selection_command
+16,9612,"/fast/home/franz.srambical/jafar/train_tokenizer.py",91,0,"",python,selection_command
+17,9746,"/fast/home/franz.srambical/jafar/train_tokenizer.py",87,0,"",python,selection_command
+18,10176,"/fast/home/franz.srambical/jafar/train_tokenizer.py",91,0,"",python,selection_command
+19,14048,"utils/train_utils.py",0,0,"import jax\nimport optax\nfrom jax.tree_util import tree_map, tree_reduce\n\n\ndef get_lr_schedule(\n    lr_schedule: str,\n    init_lr: float,\n    max_lr: float,\n    decay_end: float,\n    total_steps: int,\n    warmup_steps: int,\n    wsd_decay_steps: int,\n) -> optax.Schedule:\n    supported_schedules = [""wsd"", ""cos""]\n    if lr_schedule == ""cos"":\n        assert (\n            warmup_steps <= total_steps\n        ), ""Warmup steps can't be greater than total steps.""\n        return optax.warmup_cosine_decay_schedule(\n            init_value=init_lr,\n            peak_value=max_lr,\n            warmup_steps=warmup_steps,\n            decay_steps=total_steps,  # Note: decay_steps includes the warmup steps, so we need to pass total value\n            end_value=decay_end,\n        )\n    elif lr_schedule == ""wsd"":\n        assert (\n            warmup_steps + wsd_decay_steps <= total_steps\n        ), ""Warmup and decay period is longer than total steps.""\n        schedules = [\n            optax.linear_schedule(\n                init_value=init_lr, end_value=max_lr, transition_steps=warmup_steps\n            ),\n            optax.constant_schedule(value=max_lr),\n            optax.linear_schedule(\n                init_value=max_lr, end_value=decay_end, transition_steps=wsd_decay_steps\n            ),\n        ]\n        boundaries = [warmup_steps, total_steps - wsd_decay_steps]\n        return optax.join_schedules(schedules, boundaries)\n    else:\n        raise ValueError(\n            f""Learning rate schedule not supported. Please use one of {supported_schedules}""\n        )\n\n\ndef _count_leaf(x):\n    """"""Count parameters in a single leaf node.""""""\n    if hasattr(x, ""size""):\n        return x.size\n    return 0\n\n\ndef _count_component(component_params):\n    """"""Count total parameters in a component.""""""\n    return tree_reduce(\n        lambda x, y: x + y, tree_map(_count_leaf, component_params), initializer=0\n    )\n\n\ndef count_parameters_by_component(params):\n    """"""Count parameters for each component of the model.\n\n    Args:\n        params: Model parameters from nnx.split(model, nnx.Param, ...)\n\n    Returns:\n        Dictionary with parameter counts for each component\n    """"""\n    component_names = list(params.keys())\n    print(f""Counting all components: {component_names}"")\n\n    counts = {}\n    total_params = 0\n\n    for name in component_names:\n        component_params = params[name]\n        count = _count_component(component_params)\n        counts[name] = count\n        total_params += count\n\n    counts[""total""] = total_params\n    return counts\n\n\ndef bytes_to_gb(num_bytes):\n    return num_bytes / (1024**3)\n\n\ndef print_compiled_memory_stats(compiled_stats):\n    """"""from: https://github.com/AI-Hypercomputer/maxtext/blob/b18829fbaa48aec7ac350a03e62248e24c6a76b2/MaxText/max_utils.py#L739""""""\n    output_gb = bytes_to_gb(compiled_stats.output_size_in_bytes)\n    temp_gb = bytes_to_gb(compiled_stats.temp_size_in_bytes)\n    argument_gb = bytes_to_gb(compiled_stats.argument_size_in_bytes)\n    alias_gb = bytes_to_gb(compiled_stats.alias_size_in_bytes)\n    host_temp_gb = bytes_to_gb(compiled_stats.host_temp_size_in_bytes)\n    total_gb = output_gb + temp_gb + argument_gb - alias_gb\n    print(\n        f""Total memory size: {total_gb:.1f} GB, Output size: {output_gb:.1f} GB, Temp size: {temp_gb:.1f} GB, ""\n        f""Argument size: {argument_gb:.1f} GB, Host temp size: {host_temp_gb:.1f} GB.""\n    )\n\n\ndef print_compiled_cost_analysis(cost_stats):\n    flops = float(cost_stats.get(""flops"", 0.0))\n    bytes_accessed = float(cost_stats.get(""bytes accessed"", 0.0))\n    gb = bytes_to_gb(bytes_accessed) if bytes_accessed else 0.0\n    intensity = (flops / bytes_accessed) if bytes_accessed else float(""nan"")\n    print(\n        f""FLOPs: {flops:.3e}, Bytes: {bytes_accessed:.3e} ({gb:.1f} GB), ""\n        f""Intensity: {intensity:.1f} FLOPs/byte""\n    )\n\n\ndef print_mem_stats(label: str):\n    """"""from: https://github.com/AI-Hypercomputer/maxtext/blob/7898576359bacde81be25cb3038e348aac1f943b/MaxText/max_utils.py#L713""""""\n    print(f""\nMemstats: {label}:"")\n    try:\n        for d in jax.local_devices():\n            stats = d.memory_stats()\n            used = round(stats[""bytes_in_use""] / 2**30, 2)\n            limit = round(stats[""bytes_limit""] / 2**30, 2)\n            print(f""\tUsing (GB) {used} / {limit} ({used/limit:%}) on {d}"")\n    except (RuntimeError, KeyError, TypeError) as ex:\n        print(f""\tMemstats unavailable, error: {ex}"")\n",python,tab
+20,15843,"train_tokenizer.py",0,0,"import os\n\n# os.environ['XLA_FLAGS'] = (\n#     '--xla_python_client_mem_fraction=.98 '\n# )\n# FIXME (f.srambical): test whether this increases throughput\nos.environ['XLA_FLAGS'] = (\n    '--xla_gpu_enable_latency_hiding_scheduler=true '\n    '--xla_gpu_enable_async_collectives=true '\n)\n\nfrom dataclasses import dataclass, field\nfrom typing import cast, Optional\n\nimport einops\nimport itertools\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.tokenizer import TokenizerVQVAE\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 300_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    vq_beta: float = 0.25\n    batch_size: int = 48\n    init_lr: float = 0.0\n    max_lr: float = 3e-4\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        20000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    warmup_steps: int = 10000\n    # Tokenizer\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 1024\n    patch_size: int = 4\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.01\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_tokenizer""\n    tags: list[str] = field(default_factory=lambda: [""tokenizer""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    log_gradients: bool = False\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[TokenizerVQVAE, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        TokenizerVQVAE(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(\n    model: TokenizerVQVAE, args: Args\n) -> tuple[nnx.Optimizer, optax.Schedule]:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.Optimizer(model, tx)\n    return optimizer, lr_schedule\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.Optimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(args.data_dir, x)\n        for x in os.listdir(args.data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> ocp.CheckpointManager:\n    handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointSave,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    handler_registry.add(\n        ""dataloader_state"",\n        grain.checkpoint.CheckpointRestore,\n        cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n    )\n    checkpoint_options = ocp.CheckpointManagerOptions(\n        save_interval_steps=args.log_checkpoint_interval,\n        max_to_keep=3,\n        keep_period=args.log_checkpoint_keep_period,\n        step_format_fixed_length=6,\n        cleanup_tmp_directories=True,\n    )\n    checkpoint_manager = ocp.CheckpointManager(\n        args.ckpt_dir,\n        options=checkpoint_options,\n        handler_registry=handler_registry,\n    )\n    return checkpoint_manager\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: ocp.CheckpointManager,\n    optimizer: nnx.Optimizer,\n    grain_iterator: grain.DataLoaderIterator,\n    restore_step: Optional[int] = None,\n) -> tuple[int, nnx.Optimizer, grain.DataLoaderIterator]:\n    step = 0\n    if restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        restored = checkpoint_manager.restore(\n            restore_step,\n            args=ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                dataloader_state=grain.checkpoint.CheckpointRestore(grain_iterator),  # type: ignore\n            ),\n        )\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        grain_iterator = restored[""dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, grain_iterator\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    tokenizer, rng = build_model(args, rng)\n\n    _, params, _ = nnx.split(tokenizer, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer, lr_schedule = build_optimizer(tokenizer, args)\n    del tokenizer\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    mesh, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    grain_iterator = build_dataloader(args)\n\n    # --- Restore checkpoint ---\n    step, optimizer, grain_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, grain_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def tokenizer_loss_fn(\n        model: TokenizerVQVAE, inputs: dict\n    ) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        model.train()\n        outputs = model(inputs, training=True)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        mse = jnp.square(gt - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_clipped = gt.clip(0, 1).reshape(-1, *gt.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_clipped, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_clipped, recon)).mean()\n        _, index_counts = jnp.unique_counts(\n            jnp.ravel(outputs[""indices""]), size=args.num_latents, fill_value=0\n        )\n        codebook_usage = (index_counts != 0).mean()\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=codebook_usage,\n        )\n        return loss, (outputs[""recon""], metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.Optimizer, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        def loss_fn(model: TokenizerVQVAE) -> tuple[jax.Array, tuple[jax.Array, dict]]:\n            return tokenizer_loss_fn(model, inputs)\n\n        (loss, (recon, metrics)), grads = nnx.value_and_grad(loss_fn, has_aux=True)(\n            optimizer.model\n        )\n        optimizer.update(grads)\n        if args.log_gradients:\n            metrics[""encoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""encoder""]\n            )\n            metrics[""vq_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""vq""]\n            )\n            metrics[""decoder_gradients_std/""] = jax.tree.map(\n                lambda x: x.std(), grads[""params""][""decoder""]\n            )\n        return loss, recon, metrics\n\n    # --- TRAIN LOOP ---\n    dataloader = (\n        jax.make_array_from_process_local_data(videos_sharding, elem)\n        for elem in grain_iterator\n    )\n    if jax.process_index() == 0:\n        first_videos = next(dataloader)\n        sample_inputs = dict(videos=first_videos)\n        compiled = train_step.lower(optimizer, sample_inputs).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader = itertools.chain([first_videos], dataloader)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for videos in dataloader:\n            # --- Train step ---\n            inputs = dict(videos=videos)\n            loss, recon, metrics = train_step(optimizer, inputs)\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            metrics[""lr""] = lr_schedule(step)\n            print(f""Step {step}, loss: {loss}"")\n            step += 1\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    wandb.log(\n                        {\n                            ""loss"": loss,\n                            ""step"": step,\n                            **metrics,\n                        }\n                    )\n                if step % args.log_image_interval == 0:\n                    gt_seq = inputs[""videos""][0].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                optimizer_state = nnx.state(optimizer)\n                checkpoint_manager.save(\n                    step,\n                    args=ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            grain_iterator  # type: ignore\n                        ),\n                    ),\n                )\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+21,17040,"train_tokenizer.py",10,0,"",python,selection_command
+22,17279,"train_tokenizer.py",11,0,"",python,selection_command
+23,17319,"train_tokenizer.py",41,0,"",python,selection_command
+24,17369,"train_tokenizer.py",87,0,"",python,selection_command
+25,17389,"train_tokenizer.py",91,0,"",python,selection_command
+26,17406,"train_tokenizer.py",153,0,"",python,selection_command
+27,17436,"train_tokenizer.py",181,0,"",python,selection_command
+28,17491,"train_tokenizer.py",235,0,"",python,selection_command
+29,17521,"train_tokenizer.py",282,0,"",python,selection_command
+30,17569,"train_tokenizer.py",284,0,"",python,selection_command
+31,17591,"train_tokenizer.py",285,0,"",python,selection_command
+32,17884,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"",Log,tab
+33,22761,"train_tokenizer.py",0,0,"",python,tab
+34,22763,"TERMINAL",0,0,"",,terminal_focus
+35,23941,"TERMINAL",0,0,"source /home/franz.srambical/jafar/.venv/bin/activate",,terminal_command
+36,26286,"TERMINAL",0,0,"squeue",,terminal_command
+37,26302,"TERMINAL",0,0,"]633;C  JOBID       USER PARTITION NODES  CPUS ST         SUBMIT_TIME          START_TIME     TIME TIME_LIMIT NODELIST(REASON)\r\n]0;franz.srambical@hai-login2:~/jafar",,terminal_output
+38,28574,"TERMINAL",0,0,"salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_command
+39,28659,"TERMINAL",0,0,"]633;Csalloc: Granted job allocation 23456\r\nsalloc: Nodes hai001 are ready for job\r\n",,terminal_output
+40,29034,"TERMINAL",0,0,"Running inside SLURM, Job ID 23456.\r\n",,terminal_output
+41,29121,"TERMINAL",0,0,"]0;franz.srambical@hai-login2:~/jafar[?2004h[franz.srambical@hai001.haicore.berlin:~/jafar] $ ",,terminal_output
+42,33463,"TERMINAL",0,0,"\r(reverse-i-search)`': ",,terminal_output
+43,33751,"TERMINAL",0,0,"t': salloc --gpus=1 --ntasks-per-node=1 --cpus-per-task=1 --mem=100G",,terminal_output
+44,33926,"TERMINAL",0,0,"\ro': python3 -m MaxText.train MaxText/configs/base.yml   run_name=h100_mfu_340m   hardware=gpu   dataset_type=synthetic   steps=60   log_period=1   enable_checkpointing=False   gcs_metrics=False   metrics_file=/tmp/h100_mfu_metrics.jsonl   base_output_directory=/tmp/maxtext   attention=autoselected   per_device_batch_size=4   base_emb_dim=1536   base_num_query_heads=12   base_num_kv_heads=12   base_mlp_dim=4096   base_num_decoder_layers=10   head_dim=128   logits_via_embedding=True\rk': bash experiments/dynamics_grain_tok_restore.sh \r\n\r\r\n\r\r\n\r",,terminal_output
+45,34009,"TERMINAL",0,0,"\r[4@e': bash experiments/tokenizer_grain_checkpointing",,terminal_output
+46,34619,"TERMINAL",0,0,"\r[24@[franz.srambical@hai001.haicore.berlin:~/jafar] $ bash experiments/toke",,terminal_output
+47,34841,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+48,36266,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=23456.0 task 0: running\r\n",,terminal_output
+49,36411,"TERMINAL",0,0,"^Csrun: sending Ctrl-C to StepId=23456.0\r\nsrun: forcing job termination\r\nTraceback (most recent call last):\r\n  File ""/fast/home/franz.srambical/jafar/train_tokenizer.py"", line 19, in <module>\r\n    import optax\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/optax/__init__.py"", line 23, in <module>\r\nsrun: Job step aborted: Waiting up to 32 seconds for job step to finish.\r\n    from optax import contrib\r\n  File ""/fast/home/franz.srambical/jafar/.venv/lib/python3.10/site-packages/optax/contrib/__init__.py"", line 19, in <module>\r\n[2025-08-25T10:53:18.671] error: *** STEP 23456.0 ON hai001 CANCELLED AT 2025-08-25T10:53:18 DUE to SIGNAL Killed ***\r\n    from optax.contrib._acprop import acprop\r\n]0;franz.srambical@hai-login2:~/jafar[?2004h[franz.srambical@hai001.haicore.berlin:~/jafar] $ ",,terminal_output
+50,55421,"train_tokenizer.py",0,0,"Switched from branch 'tokenizer-fwd-half-precision' to 'simplified-param-calculation'",python,git_branch_checkout
+51,60005,"TERMINAL",0,0,"bash experiments/tokenizer_grain_checkpointing.sh ",,terminal_output
+52,60162,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+53,65422,"TERMINAL",0,0,"2025-08-25 10:53:47.689852: F external/xla/xla/parse_flags_from_env.cc:234] Unknown flag in XLA_FLAGS: --xla_gpu_enable_async_collectives=true\r\n",,terminal_output
+54,70894,"TERMINAL",0,0,"srun: error: hai001: task 0: Aborted (core dumped)\r\n]0;franz.srambical@hai-login2:~/jafar[?2004h[franz.srambical@hai001.haicore.berlin:~/jafar] $ ",,terminal_output
+55,71219,"TERMINAL",0,0,"^C[?2004l\r[?2004h[?2004l\r\r\n]0;franz.srambical@hai-login2:~/jafar[?2004h[franz.srambical@hai001.haicore.berlin:~/jafar] $ ",,terminal_output
+56,72219,"train_tokenizer.py",284,0,"",python,selection_command
+57,72346,"train_tokenizer.py",282,0,"",python,selection_command
+58,72499,"train_tokenizer.py",235,0,"",python,selection_command
+59,72659,"train_tokenizer.py",181,0,"",python,selection_command
+60,72799,"train_tokenizer.py",153,0,"",python,selection_command
+61,73289,"train_tokenizer.py",153,27,"os.environ['XLA_FLAGS'] = (",python,selection_command
+62,73379,"train_tokenizer.py",153,0,"",python,selection_command
+63,73554,"train_tokenizer.py",181,0,"",python,selection_command
+64,74004,"train_tokenizer.py",153,0,"",python,selection_command
+65,74219,"train_tokenizer.py",153,27,"os.environ['XLA_FLAGS'] = (",python,selection_command
+66,74299,"train_tokenizer.py",153,81,"os.environ['XLA_FLAGS'] = (\n    '--xla_gpu_enable_latency_hiding_scheduler=true '",python,selection_command
+67,74444,"train_tokenizer.py",153,128,"os.environ['XLA_FLAGS'] = (\n    '--xla_gpu_enable_latency_hiding_scheduler=true '\n    '--xla_gpu_enable_async_collectives=true '",python,selection_command
+68,74599,"train_tokenizer.py",153,130,"os.environ['XLA_FLAGS'] = (\n    '--xla_gpu_enable_latency_hiding_scheduler=true '\n    '--xla_gpu_enable_async_collectives=true '\n)",python,selection_command
+69,76005,"train_tokenizer.py",282,0,"# ",python,content
+70,76005,"train_tokenizer.py",235,0,"# ",python,content
+71,76005,"train_tokenizer.py",181,0,"# ",python,content
+72,76005,"train_tokenizer.py",153,0,"# ",python,content
+73,76272,"train_tokenizer.py",288,0,"",python,selection_command
+74,77492,"TERMINAL",0,0,"bash experiments/tokenizer_grain_checkpointing.sh ",,terminal_output
+75,77831,"TERMINAL",0,0,"\r\n[?2004l\r",,terminal_output
+76,82227,"TERMINAL",0,0,"Running on 1 devices.\r\n",,terminal_output
+77,86097,"TERMINAL",0,0,"Counting all components: ['decoder', 'encoder', 'vq']\r\nParameter counts:\r\n{'decoder': 16858736, 'encoder': 16858752, 'vq': 32768, 'total': 33750256}\r\n",,terminal_output
+78,117375,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=23456.2 task 0: running\r\n",,terminal_output
+79,126879,"TERMINAL",0,0,"2025-08-25 10:54:48.958239: E external/xla/xla/stream_executor/cuda/cuda_timer.cc:86] Delay kernel timed out: measured time has sub-optimal accuracy. There may be a missing warmup execution, please investigate in Nsight Systems.\r\n",,terminal_output
+80,126979,"TERMINAL",0,0,"2025-08-25 10:54:49.236371: E external/xla/xla/stream_executor/cuda/cuda_timer.cc:86] Delay kernel timed out: measured time has sub-optimal accuracy. There may be a missing warmup execution, please investigate in Nsight Systems.\r\n",,terminal_output
+81,136850,"TERMINAL",0,0,"Total memory size: 9.5 GB, Output size: 0.4 GB, Temp size: 9.0 GB, Argument size: 0.4 GB, Host temp size: 0.0 GB.\r\nFLOPs: 3.306e+11, Bytes: 2.483e+11 (231.2 GB), Intensity: 1.3 FLOPs/byte\r\nStarting training from step 0...\r\n",,terminal_output
+82,137039,"TERMINAL",0,0,"\r\nMemstats: After params initialized:\r\n\tUsing (GB) 0.44 / 59.39 (0.740865%) on cuda:0\r\n",,terminal_output
+83,137484,"TERMINAL",0,0,"Step 0, loss: 0.25431719422340393\r\n",,terminal_output
+84,152028,"TERMINAL",0,0,"Step 1, loss: 0.2657988965511322\r\n",,terminal_output
+85,152324,"TERMINAL",0,0,"Step 2, loss: 0.27664220333099365\r\n",,terminal_output
+86,154001,"TERMINAL",0,0,"Step 3, loss: 0.2549745738506317\r\n",,terminal_output
+87,154002,"TERMINAL",0,0,"Step 4, loss: 0.23929359018802643\r\nSaved checkpoint at step 5\r\n",,terminal_output
+88,154058,"TERMINAL",0,0,"Step 5, loss: 0.22147467732429504\r\n",,terminal_output
+89,154576,"TERMINAL",0,0,"Step 6, loss: 0.21243974566459656\r\n",,terminal_output
+90,155068,"TERMINAL",0,0,"Step 7, loss: 0.20390529930591583\r\n",,terminal_output
+91,155888,"TERMINAL",0,0,"Step 8, loss: 0.2014799565076828\r\n",,terminal_output
+92,156305,"TERMINAL",0,0,"Step 9, loss: 0.19134578108787537\r\nSaved checkpoint at step 10\r\n",,terminal_output
+93,156697,"TERMINAL",0,0,"Step 10, loss: 0.1849638819694519\r\n",,terminal_output
+94,157169,"TERMINAL",0,0,"Step 11, loss: 0.18320240080356598\r\n",,terminal_output
+95,157528,"TERMINAL",0,0,"Step 12, loss: 0.17448516190052032\r\n",,terminal_output
+96,157928,"TERMINAL",0,0,"Step 13, loss: 0.1701081246137619\r\n",,terminal_output
+97,158527,"TERMINAL",0,0,"Step 14, loss: 0.16334110498428345\r\nSaved checkpoint at step 15\r\n",,terminal_output
+98,159169,"TERMINAL",0,0,"Step 15, loss: 0.1604066789150238\r\n",,terminal_output
+99,159681,"TERMINAL",0,0,"Step 16, loss: 0.15722087025642395\r\n",,terminal_output
+100,160269,"TERMINAL",0,0,"Step 17, loss: 0.15154121816158295\r\n",,terminal_output
+101,160769,"TERMINAL",0,0,"Step 18, loss: 0.14370152354240417\r\n",,terminal_output
+102,161329,"TERMINAL",0,0,"Step 19, loss: 0.13980640470981598\r\nSaved checkpoint at step 20\r\n",,terminal_output
+103,162026,"TERMINAL",0,0,"Step 20, loss: 0.13828369975090027\r\n",,terminal_output
+104,162689,"TERMINAL",0,0,"Step 21, loss: 0.13768532872200012\r\n",,terminal_output
+105,163123,"TERMINAL",0,0,"Step 22, loss: 0.13598142564296722\r\n",,terminal_output
+106,163964,"TERMINAL",0,0,"Step 23, loss: 0.13169215619564056\r\n",,terminal_output
+107,164595,"TERMINAL",0,0,"Step 24, loss: 0.12741072475910187\r\nSaved checkpoint at step 25\r\n",,terminal_output
+108,165523,"TERMINAL",0,0,"Step 25, loss: 0.12643606960773468\r\n",,terminal_output
+109,166271,"TERMINAL",0,0,"Step 26, loss: 0.1263691782951355\r\n",,terminal_output
+110,166866,"TERMINAL",0,0,"Step 27, loss: 0.12507067620754242\r\n",,terminal_output
+111,167584,"TERMINAL",0,0,"Step 28, loss: 0.1210612878203392\r\n",,terminal_output
+112,168160,"TERMINAL",0,0,"Step 29, loss: 0.11947143077850342\r\nSaved checkpoint at step 30\r\n",,terminal_output
+113,168656,"TERMINAL",0,0,"Step 30, loss: 0.11961992084980011\r\n",,terminal_output
+114,169241,"TERMINAL",0,0,"Step 31, loss: 0.11844073981046677\r\n",,terminal_output
+115,170428,"TERMINAL",0,0,"Step 32, loss: 0.11614794284105301\r\n",,terminal_output
+116,170429,"TERMINAL",0,0,"Step 33, loss: 0.11374820023775101\r\n",,terminal_output
+117,171149,"TERMINAL",0,0,"^Csrun: interrupt (one more within 1 sec to abort)\r\nsrun: StepId=23456.2 task 0: running\r\n^Csrun: sending Ctrl-C to StepId=23456.2\r\nsrun: forcing job termination\r\nsrun: Job step aborted: Waiting up to 32 seconds for job step to finish.\r\n[2025-08-25T10:55:33.214] error: *** STEP 23456.2 ON hai001 CANCELLED AT 2025-08-25T10:55:33 DUE to SIGNAL Killed ***\r\n",,terminal_output
+118,171645,"TERMINAL",0,0,"]0;franz.srambical@hai-login2:~/jafar[?2004h[franz.srambical@hai001.haicore.berlin:~/jafar] $ ^C[?2004l\r[?2004h[?2004l\r\r\n]0;franz.srambical@hai-login2:~/jafar[?2004h[franz.srambical@hai001.haicore.berlin:~/jafar] $ ",,terminal_output

1f15334ab7e6820c9fda17c961659882ef9853cc80f7356b9a9b22f286fd7389/crowd-code-eb393a0e-6aef-412a-bef0-408ff847ad2c1764928526101-2025_12_05-10.55.35.367/source.csv

@@ -0,0 +1,2 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,503,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"10:55:35 AM [info] Activating crowd-code\n10:55:35 AM [info] Recording started\n10:55:35 AM [info] Initializing git provider using file system watchers...\n10:55:35 AM [info] Git repository found\n10:55:35 AM [info] Git provider initialized successfully\n10:55:35 AM [info] Initial git state: [object Object]\n",Log,tab

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-02f1e1f4-094a-4d7b-957e-88354d5de1001754231283463-2025_08_03-16.28.12.317/source.csv

@@ -0,0 +1,515 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,690,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"4:28:12 PM [info] Activating crowd-code\n4:28:12 PM [info] Recording started\n4:28:12 PM [info] Initializing git provider using file system watchers...\n4:28:12 PM [info] Git repository found\n4:28:12 PM [info] Git provider initialized successfully\n",Log,tab
+3,853,"extension-output-pdoom-org.crowd-code-#1-crowd-code",245,0,"4:28:13 PM [info] Initial git state: [object Object]\n",Log,content
+4,4654,"TERMINAL",0,0,"queue",,terminal_command
+5,4724,"TERMINAL",0,0,"]633;E;2025-08-03 16:28:16 queue;1fa2723a-e415-422b-a7e2-4358520da61a]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1991.localdomain: Sun Aug  3 16:28:16 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3388151 accelerat train_to tum_cte0  R    7:25:27\t  2 hkn[0533-0534]3388153 accelerat train_to tum_cte0  R    7:25:27\t  2 hkn[0717-0718]3393066 accelerat train_dy tum_cte0  R    4:10:10     12 hkn[0701-0712]3393065 accelerat train_dy tum_cte0  R    4:12:20\t  3 hkn[0808-0810]",,terminal_output
+6,5767,"TERMINAL",0,0,"bash",,terminal_focus
+7,5768,"TERMINAL",0,0,"78811",,terminal_output
+8,6238,"TERMINAL",0,0,"watch",,terminal_focus
+9,6804,"TERMINAL",0,0,"8303033",,terminal_output
+10,7849,"TERMINAL",0,0,"201144",,terminal_output
+11,8886,"TERMINAL",0,0,"12255",,terminal_output
+12,9718,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0",,terminal_output
+13,39964,"TERMINAL",0,0,"source ../../jafar/.venv/bin/activate",,terminal_command
+14,40007,"TERMINAL",0,0,"]633;E;2025-08-03 16:28:52 source ../../jafar/.venv/bin/activate;1fa2723a-e415-422b-a7e2-4358520da61a]633;C]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0",,terminal_output
+15,45211,"TERMINAL",0,0,"bash",,terminal_focus
+16,46116,"TERMINAL",0,0,"../../jafar/.venv/bin/activate",,terminal_command
+17,50378,"TERMINAL",0,0,"source ../../jafar/.venv/bin/activate",,terminal_command
+18,51001,"TERMINAL",0,0,"bash",,terminal_focus
+19,53162,"sample.py",0,0,"from dataclasses import dataclass\nimport time\n\nimport dm_pix as pix\nimport einops\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\nimport numpy as np\nfrom orbax.checkpoint import PyTreeCheckpointer\nfrom PIL import Image, ImageDraw\nimport tyro\n\nfrom genie import Genie\nfrom utils.dataloader import get_dataloader\n\n\n@dataclass\nclass Args:\n    # Experiment\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_resolution: int = 64\n    data_dir: str = ""data/coinrun_episodes""\n    checkpoint: str = """"\n    # Sampling\n    batch_size: int = 1\n    maskgit_steps: int = 25\n    temperature: float = 1.0\n    sample_argmax: bool = True\n    start_frame: int = 0\n    # Tokenizer checkpoint\n    tokenizer_dim: int = 512\n    latent_patch_dim: int = 32\n    num_patch_latents: int = 1024\n    patch_size: int = 4\n    tokenizer_num_blocks: int = 8\n    tokenizer_num_heads: int = 8\n    # LAM checkpoint\n    lam_dim: int = 512\n    latent_action_dim: int = 32\n    num_latent_actions: int = 6\n    lam_patch_size: int = 16\n    lam_num_blocks: int = 8\n    lam_num_heads: int = 8\n    # Dynamics checkpoint\n    dyna_dim: int = 512\n    dyna_num_blocks: int = 12\n    dyna_num_heads: int = 8\n\n\nargs = tyro.cli(Args)\nrng = jax.random.PRNGKey(args.seed)\n\n# --- Load Genie checkpoint ---\ngenie = Genie(\n    # Tokenizer\n    in_dim=args.image_channels,\n    tokenizer_dim=args.tokenizer_dim,\n    latent_patch_dim=args.latent_patch_dim,\n    num_patch_latents=args.num_patch_latents,\n    patch_size=args.patch_size,\n    tokenizer_num_blocks=args.tokenizer_num_blocks,\n    tokenizer_num_heads=args.tokenizer_num_heads,\n    # LAM\n    lam_dim=args.lam_dim,\n    latent_action_dim=args.latent_action_dim,\n    num_latent_actions=args.num_latent_actions,\n    lam_patch_size=args.lam_patch_size,\n    lam_num_blocks=args.lam_num_blocks,\n    lam_num_heads=args.lam_num_heads,\n    # Dynamics\n    dyna_dim=args.dyna_dim,\n    dyna_num_blocks=args.dyna_num_blocks,\n    dyna_num_heads=args.dyna_num_heads,\n)\nrng, _rng = jax.random.split(rng)\nimage_shape = (args.image_resolution, args.image_resolution, args.image_channels)\ndummy_inputs = dict(\n    videos=jnp.zeros((args.batch_size, args.seq_len, *image_shape), dtype=jnp.float32),\n    mask_rng=_rng,\n)\nrng, _rng = jax.random.split(rng)\nparams = genie.init(_rng, dummy_inputs)\nckpt = PyTreeCheckpointer().restore(args.checkpoint)[""model""][""params""][""params""]\nparams[""params""].update(ckpt)\n\n\ndef _sampling_wrapper(module, batch):\n    return module.sample(batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax)\n\n# --- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid\n\n\n# --- Get video + latent actions ---\ndataloader = get_dataloader(args.data_dir, args.seq_len, args.batch_size)\nvideo_batch = next(iter(dataloader))\n# Get latent actions from first video only\nfirst_video = video_batch[:1]\nbatch = dict(videos=first_video)\naction_batch = genie.apply(params, batch, False, method=Genie.vq_encode)\naction_batch = action_batch.reshape(1, args.seq_len - 1, 1)\n# Use actions from first video for all videos\naction_batch = jnp.repeat(action_batch, video_batch.shape[0], axis=0)\n\n# --- Sample + evaluate video ---\nvid = _autoreg_sample(rng, video_batch, action_batch)\ngt = video_batch[:, : vid.shape[1]].clip(0, 1).reshape(-1, *video_batch.shape[2:])\nrecon = vid.clip(0, 1).reshape(-1, *vid.shape[2:])\nssim = pix.ssim(gt[:, args.start_frame + 1 :], recon[:, args.start_frame + 1 :]).mean()\nprint(f""SSIM: {ssim}"")\n\n# --- Construct video ---\nfirst_true = (video_batch[0:1] * 255).astype(np.uint8)\nfirst_pred = (vid[0:1] * 255).astype(np.uint8)\nfirst_video_comparison = np.zeros((2, *vid.shape[1:5]), dtype=np.uint8)\nfirst_video_comparison[0] = first_true[:, : vid.shape[1]]\nfirst_video_comparison[1] = first_pred\n# For other videos, only show generated video\nother_preds = (vid[1:] * 255).astype(np.uint8)\nall_frames = np.concatenate([first_video_comparison, other_preds], axis=0)\nflat_vid = einops.rearrange(all_frames, ""n t h w c -> t h (n w) c"")\n\n# --- Save video ---\nimgs = [Image.fromarray(img) for img in flat_vid]\n# Write actions on each frame\nfor img, action in zip(imgs[1:], action_batch[0, :, 0]):\n    d = ImageDraw.Draw(img)\n    d.text((2, 2), f""{action}"", fill=255)\nimgs[0].save(\n    f""generation_{time.time()}.gif"",\n    save_all=True,\n    append_images=imgs[1:],\n    duration=250,\n    loop=0,\n)\n",python,tab
+20,57765,"genie.py",0,0,"from typing import Dict, Any\n\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\n\nfrom models.dynamics import DynamicsMaskGIT\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\n\nclass Genie(nn.Module):\n    """"""Genie model""""""\n\n    # --- Tokenizer ---\n    in_dim: int\n    tokenizer_dim: int\n    latent_patch_dim: int\n    num_patch_latents: int\n    patch_size: int\n    tokenizer_num_blocks: int\n    tokenizer_num_heads: int\n    # --- LAM ---\n    lam_dim: int\n    latent_action_dim: int\n    num_latent_actions: int\n    lam_patch_size: int\n    lam_num_blocks: int\n    lam_num_heads: int\n    # --- Dynamics ---\n    dyna_dim: int\n    dyna_num_blocks: int\n    dyna_num_heads: int\n    dropout: float = 0.0\n    mask_limit: float = 0.0\n\n    def setup(self):\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.dynamics = DynamicsMaskGIT(\n            model_dim=self.dyna_dim,\n            num_latents=self.num_patch_latents,\n            num_blocks=self.dyna_num_blocks,\n            num_heads=self.dyna_num_heads,\n            dropout=self.dropout,\n            mask_limit=self.mask_limit,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        tokenizer_outputs = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        lam_outputs = self.lam.vq_encode(batch[""videos""], training=False)\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(tokenizer_outputs[""indices""]),\n            latent_actions=jax.lax.stop_gradient(lam_outputs[""z_q""]),\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_outputs = self.dynamics(outputs, training)\n        outputs.update(dyna_outputs)\n        mle_indices = jnp.argmax(outputs[""token_logits""], axis=-1)\n        outputs[""recon""] = self.tokenizer.decode(\n            mle_indices, batch[""videos""].shape[2:4]\n        )\n        return outputs\n    \n\n    @nn.compact\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        steps: int = 25,\n        temperature: int = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]\n        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # --- Initialize MaskGIT ---\n        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n        init_carry = (\n            batch[""rng""],\n            new_frame_idxs,\n            init_mask,\n            token_idxs,\n            action_tokens,\n        )\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n\n        # --- Run MaskGIT loop ---\n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n        new_frame_idxs = final_carry[1]\n        new_frame_pixels = self.tokenizer.decode(\n            jnp.expand_dims(new_frame_idxs, 1),\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return new_frame_pixels\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n\ndef restore_genie_components(params: Dict[str, Any], tokenizer: str, lam: str):\n    """"""Restore pre-trained Genie components""""""\n    params[""params""][""tokenizer""].update(\n        PyTreeCheckpointer().restore(tokenizer)[""model""][""params""][""params""]\n    )\n    params[""params""][""lam""].update(\n        PyTreeCheckpointer().restore(lam)[""model""][""params""][""params""]\n    )\n    return params\n",python,tab
+21,60432,"genie.py",545,0,"",python,selection_mouse
+22,60433,"genie.py",544,0,"",python,selection_command
+23,93412,"genie.py",431,0,"",python,selection_mouse
+24,93414,"genie.py",430,0,"",python,selection_command
+25,96478,"TERMINAL",0,0,"bash",,terminal_focus
+26,123732,"genie.py",315,0,"",python,selection_mouse
+27,128367,"genie.py",136,0,"",python,selection_mouse
+28,128373,"genie.py",135,0,"",python,selection_command
+29,132808,"genie.py",136,0,"\n",python,content
+30,132976,"genie.py",137,0,"i",python,content
+31,132977,"genie.py",138,0,"",python,selection_keyboard
+32,133069,"genie.py",138,0,"m",python,content
+33,133071,"genie.py",139,0,"",python,selection_keyboard
+34,133288,"genie.py",139,0,"p",python,content
+35,133290,"genie.py",140,0,"",python,selection_keyboard
+36,133346,"genie.py",140,0,"o",python,content
+37,133347,"genie.py",141,0,"",python,selection_keyboard
+38,133590,"genie.py",141,0,"r",python,content
+39,133592,"genie.py",142,0,"",python,selection_keyboard
+40,133785,"genie.py",142,0,"t",python,content
+41,133787,"genie.py",143,0,"",python,selection_keyboard
+42,133882,"genie.py",143,0," ",python,content
+43,133885,"genie.py",144,0,"",python,selection_keyboard
+44,134211,"genie.py",144,0,"e",python,content
+45,134212,"genie.py",145,0,"",python,selection_keyboard
+46,134282,"genie.py",145,0,"i",python,content
+47,134283,"genie.py",146,0,"",python,selection_keyboard
+48,134352,"genie.py",146,0,"n",python,content
+49,134355,"genie.py",147,0,"",python,selection_keyboard
+50,134746,"genie.py",147,0,"o",python,content
+51,134748,"genie.py",148,0,"",python,selection_keyboard
+52,134896,"genie.py",148,0,"p",python,content
+53,134898,"genie.py",149,0,"",python,selection_keyboard
+54,134990,"genie.py",149,0,"s",python,content
+55,134992,"genie.py",150,0,"",python,selection_keyboard
+56,135753,"genie.py",149,0,"",python,selection_command
+57,153495,"genie.py",6055,0,"",python,selection_mouse
+58,153498,"genie.py",6054,0,"",python,selection_command
+59,165459,"genie.py",6331,1,">",python,selection_command
+60,169297,"genie.py",6331,0,"",python,selection_command
+61,179994,"TERMINAL",0,0,"git branch",,terminal_command
+62,180037,"TERMINAL",0,0,"]633;E;2025-08-03 16:31:12 git branch;d4f80100-2225-40b3-b029-72e571a8ef11]633;C[?1h=\r* genie-sampling\r\n  main\r\n\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0",,terminal_output
+63,180789,"genie.py",3132,0,"",python,selection_mouse
+64,182029,"genie.py",3099,0,"",python,selection_mouse
+65,182080,"genie.py",3098,0,"",python,selection_command
+66,182911,"genie.py",3109,0,"",python,selection_mouse
+67,183049,"genie.py",3108,14,"new_frame_idxs",python,selection_mouse
+68,186845,"genie.py",3356,0,"",python,selection_mouse
+69,193102,"TERMINAL",0,0,"bash",,terminal_focus
+70,194098,"genie.py",0,0,"",python,tab
+71,194801,"TERMINAL",0,0,"bash",,terminal_focus
+72,196629,"TERMINAL",0,0,"git diff",,terminal_command
+73,196687,"TERMINAL",0,0,"]633;E;2025-08-03 16:31:28 git diff;d4f80100-2225-40b3-b029-72e571a8ef11]633;C[?1h=\rdiff --git a/genie.py b/genie.py\r\nindex 50729ab..7314771 100644\r\n--- a/genie.py\r\n+++ b/genie.py\r\n@@ -4,6 +4,7 @@ from orbax.checkpoint import PyTreeCheckpointer\r\n import jax\r\n import jax.numpy as jnp\r\n import flax.linen as nn\r\n+import einops\r\n \r\n from models.dynamics import DynamicsMaskGIT\r\n from models.lam import LatentActionModel\r\n:",,terminal_output
+74,198810,"genie.py",5699,0,"",python,selection_mouse
+75,198812,"genie.py",5698,0,"",python,selection_command
+76,200328,"genie.py",5769,0,"",python,selection_mouse
+77,213897,"genie.py",3703,0,"",python,selection_mouse
+78,213900,"genie.py",3702,0,"",python,selection_command
+79,215629,"genie.py",3739,1,"T",python,selection_command
+80,215683,"genie.py",3829,2,"to",python,selection_command
+81,215757,"genie.py",3829,3,"tok",python,selection_command
+82,215913,"genie.py",3829,4,"toke",python,selection_command
+83,215999,"genie.py",3829,5,"token",python,selection_command
+84,216606,"genie.py",4677,6,"token_",python,selection_command
+85,217072,"genie.py",4677,7,"token_i",python,selection_command
+86,217219,"genie.py",4677,8,"token_id",python,selection_command
+87,217663,"genie.py",4677,9,"token_idx",python,selection_command
+88,218218,"genie.py",4677,10,"token_idxs",python,selection_command
+89,237629,"genie.py",3116,0,"",python,selection_mouse
+90,237773,"genie.py",3108,14,"new_frame_idxs",python,selection_mouse
+91,356558,"genie.py",6658,0,"",python,selection_mouse
+92,364385,"genie.py",4471,2187,"\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+93,364399,"genie.py",4360,2298,"      lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+94,364414,"genie.py",4323,2335,"       # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+95,364440,"genie.py",4262,2396,"    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+96,364530,"genie.py",4261,2397,"\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+97,364607,"genie.py",2761,3897,"    def sample(\n        self,\n        batch: Dict[str, Any],\n        steps: int = 25,\n        temperature: int = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]\n        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # --- Initialize MaskGIT ---\n        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n        init_carry = (\n            batch[""rng""],\n            new_frame_idxs,\n            init_mask,\n            token_idxs,\n            action_tokens,\n        )\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n\n        # --- Run MaskGIT loop ---\n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n        new_frame_idxs = final_carry[1]\n        new_frame_pixels = self.tokenizer.decode(\n            jnp.expand_dims(new_frame_idxs, 1),\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return new_frame_pixels\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+98,364700,"genie.py",2745,3913,"    @nn.compact\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        steps: int = 25,\n        temperature: int = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]\n        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # --- Initialize MaskGIT ---\n        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n        init_carry = (\n            batch[""rng""],\n            new_frame_idxs,\n            init_mask,\n            token_idxs,\n            action_tokens,\n        )\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n\n        # --- Run MaskGIT loop ---\n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n        new_frame_idxs = final_carry[1]\n        new_frame_pixels = self.tokenizer.decode(\n            jnp.expand_dims(new_frame_idxs, 1),\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return new_frame_pixels\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n",python,selection_mouse
+99,365586,"genie.py",2745,3914,"",python,content
+100,367129,"genie.py",2743,0,"",python,selection_mouse
+101,368150,"genie.py",2744,0,"",python,selection_mouse
+102,369159,"genie.py",2744,0,"    @nn.compact\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        seq_len: int,\n        steps: int = 25,\n        temperature: float = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        """"""\n        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n\n        - Input frames are tokenized once.\n        - Future frames are generated autoregressively in token space.\n        - All frames are detokenized in a single pass.\n\n        Note:\n        - For interactive or step-wise sampling, detokenization should occur after each action.\n        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n        - Temporal causal structure is preserved by \n            a) reapplying the mask before each decoding step.\n            b) a temporal causal mask is applied within each ST-transformer block.\n\n        Dimension keys:\n            B: batch size  \n            T: number of input (conditioning) frames  \n            N: patches per frame  \n            S: sequence length  \n            A: action space  \n            D: model latent dimension\n        """"""\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""] # (B, T, N)\n        B, T, N = token_idxs.shape\n        pad_shape = (B, seq_len - T, N)\n        pad = jnp.zeros(pad_shape, dtype=token_idxs.dtype)\n        token_idxs = jnp.concatenate([token_idxs, pad], axis=1) # (B, S, N)\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n    \n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n\n        def generation_step_fn(carry, step_t):\n            rng, current_token_idxs = carry\n            rng, step_rng = jax.random.split(rng)\n\n            # Mask current and future frames (i.e., t >= step_t)\n            mask = jnp.arange(seq_len) >= step_t # (S,)\n            mask = jnp.broadcast_to(mask[None, :, None], (B, seq_len, N)) # (B, S, N)\n            mask = mask.astype(bool)\n            masked_token_idxs = current_token_idxs * ~mask\n\n            # --- Initialize and run MaskGIT loop ---\n            init_carry_maskgit = (\n                step_rng,\n                masked_token_idxs,\n                mask,\n                action_tokens,\n            )\n            final_carry_maskgit, _ = loop_fn(init_carry_maskgit, jnp.arange(steps))\n            updated_token_idxs = final_carry_maskgit[1]\n            new_carry = (rng, updated_token_idxs)\n            return new_carry, None\n\n        # --- Run the autoregressive generation using scan ---\n        initial_carry = (batch[""rng""], token_idxs)\n        timesteps_to_scan = jnp.arange(T, seq_len)\n        final_carry, _ = jax.lax.scan(\n            generation_step_fn,\n            initial_carry,\n            timesteps_to_scan\n        )\n        final_token_idxs = final_carry[1]\n\n        # --- Decode all tokens at once at the end ---\n        final_frames = self.tokenizer.decode(\n            final_token_idxs,\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return final_frames\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None",python,content
+103,370006,"genie.py",8679,0,"\n        ",python,content
+104,370116,"genie.py",8684,4,"",python,content
+105,371031,"genie.py",8684,0,"\n    ",python,content
+106,371365,"genie.py",8685,4,"",python,content
+107,377778,"genie.py",0,0,"",python,tab
+108,377844,"genie.py",137,0,"",python,selection_command
+109,379545,"genie.py",0,0,"from typing import Dict, Any\n\nfrom orbax.checkpoint import PyTreeCheckpointer\nimport jax\nimport jax.numpy as jnp\nimport flax.linen as nn\n\nfrom models.dynamics import DynamicsMaskGIT\nfrom models.lam import LatentActionModel\nfrom models.tokenizer import TokenizerVQVAE\n\n\nclass Genie(nn.Module):\n    """"""Genie model""""""\n\n    # --- Tokenizer ---\n    in_dim: int\n    tokenizer_dim: int\n    latent_patch_dim: int\n    num_patch_latents: int\n    patch_size: int\n    tokenizer_num_blocks: int\n    tokenizer_num_heads: int\n    # --- LAM ---\n    lam_dim: int\n    latent_action_dim: int\n    num_latent_actions: int\n    lam_patch_size: int\n    lam_num_blocks: int\n    lam_num_heads: int\n    # --- Dynamics ---\n    dyna_dim: int\n    dyna_num_blocks: int\n    dyna_num_heads: int\n    dropout: float = 0.0\n    mask_limit: float = 0.0\n\n    def setup(self):\n        self.tokenizer = TokenizerVQVAE(\n            in_dim=self.in_dim,\n            model_dim=self.tokenizer_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_patch_latents,\n            patch_size=self.patch_size,\n            num_blocks=self.tokenizer_num_blocks,\n            num_heads=self.tokenizer_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.lam = LatentActionModel(\n            in_dim=self.in_dim,\n            model_dim=self.lam_dim,\n            latent_dim=self.latent_patch_dim,\n            num_latents=self.num_latent_actions,\n            patch_size=self.lam_patch_size,\n            num_blocks=self.lam_num_blocks,\n            num_heads=self.lam_num_heads,\n            dropout=0.0,\n            codebook_dropout=0.0,\n        )\n        self.dynamics = DynamicsMaskGIT(\n            model_dim=self.dyna_dim,\n            num_latents=self.num_patch_latents,\n            num_blocks=self.dyna_num_blocks,\n            num_heads=self.dyna_num_heads,\n            dropout=self.dropout,\n            mask_limit=self.mask_limit,\n        )\n\n    def __call__(self, batch: Dict[str, Any], training: bool = True) -> Dict[str, Any]:\n        tokenizer_outputs = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        lam_outputs = self.lam.vq_encode(batch[""videos""], training=False)\n        outputs = dict(\n            video_tokens=jax.lax.stop_gradient(tokenizer_outputs[""indices""]),\n            latent_actions=jax.lax.stop_gradient(lam_outputs[""z_q""]),\n        )\n        outputs[""mask_rng""] = batch[""mask_rng""]\n        dyna_outputs = self.dynamics(outputs, training)\n        outputs.update(dyna_outputs)\n        mle_indices = jnp.argmax(outputs[""token_logits""], axis=-1)\n        outputs[""recon""] = self.tokenizer.decode(\n            mle_indices, batch[""videos""].shape[2:4]\n        )\n        return outputs\n\n    @nn.compact\n    def sample(\n        self,\n        batch: Dict[str, Any],\n        steps: int = 25,\n        temperature: int = 1,\n        sample_argmax: bool = False,\n    ) -> Any:\n        # --- Encode videos and actions ---\n        tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n        token_idxs = tokenizer_out[""indices""]\n        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n        action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n\n        # --- Initialize MaskGIT ---\n        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n        init_carry = (\n            batch[""rng""],\n            new_frame_idxs,\n            init_mask,\n            token_idxs,\n            action_tokens,\n        )\n        MaskGITLoop = nn.scan(\n            MaskGITStep,\n            variable_broadcast=""params"",\n            split_rngs={""params"": False},\n            in_axes=0,\n            out_axes=0,\n            length=steps,\n        )\n\n        # --- Run MaskGIT loop ---\n        loop_fn = MaskGITLoop(\n            dynamics=self.dynamics,\n            tokenizer=self.tokenizer,\n            temperature=temperature,\n            sample_argmax=sample_argmax,\n            steps=steps,\n        )\n        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n        new_frame_idxs = final_carry[1]\n        new_frame_pixels = self.tokenizer.decode(\n            jnp.expand_dims(new_frame_idxs, 1),\n            video_hw=batch[""videos""].shape[2:4],\n        )\n        return new_frame_pixels\n\n    def vq_encode(self, batch, training) -> Dict[str, Any]:\n        # --- Preprocess videos ---\n        lam_output = self.lam.vq_encode(batch[""videos""], training=training)\n        return lam_output[""indices""]\n\n\nclass MaskGITStep(nn.Module):\n    dynamics: nn.Module\n    tokenizer: nn.Module\n    temperature: float\n    sample_argmax: bool\n    steps: int\n\n    @nn.compact\n    def __call__(self, carry, x):\n        rng, token_idxs, mask, action_tokens = carry\n        step = x\n        N = token_idxs.shape[2]\n\n        # --- Construct + encode video ---\n        vid_embed = self.dynamics.patch_embed(token_idxs) # (B, S, N, D)\n        mask_token = self.dynamics.mask_token  # (1, 1, 1, D,)\n        mask_expanded = mask[..., None] # (B, S, N, 1) \n        vid_embed = jnp.where(mask_expanded, mask_token, vid_embed)\n\n        # --- Predict transition ---\n        act_embed = self.dynamics.action_up(action_tokens)\n        vid_embed += jnp.pad(act_embed, ((0, 0), (1, 0), (0, 0), (0, 0)))\n        unmasked_ratio = jnp.cos(jnp.pi * (step + 1) / (self.steps * 2))\n        step_temp = self.temperature * (1.0 - unmasked_ratio)\n        final_logits = self.dynamics.dynamics(vid_embed) / step_temp\n\n        # --- Sample new tokens for final frame ---\n        if self.sample_argmax:\n            sampled_token_idxs = jnp.argmax(final_logits, axis=-1)\n        else:\n            rng, _rng = jax.random.split(rng)\n            sampled_token_idxs = jnp.where(\n                step == self.steps - 1,\n                jnp.argmax(final_logits, axis=-1),\n                jax.random.categorical(_rng, final_logits),\n            )\n        gather_fn = jax.vmap(jax.vmap(jax.vmap(lambda x, y: x[y])))\n        final_token_probs = gather_fn(jax.nn.softmax(final_logits), sampled_token_idxs)\n        final_token_probs += ~mask\n        # Update masked tokens only\n        token_idxs = jnp.where(mask, sampled_token_idxs, token_idxs)\n\n        # --- Update mask ---\n        num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n        mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n        new_mask = mask_update_fn(mask, sorted_idxs)\n\n        new_carry = (rng, token_idxs, new_mask, action_tokens)\n        return new_carry, None\n\n\ndef restore_genie_components(params: Dict[str, Any], tokenizer: str, lam: str):\n    """"""Restore pre-trained Genie components""""""\n    params[""params""][""tokenizer""].update(\n        PyTreeCheckpointer().restore(tokenizer)[""model""][""params""][""params""]\n    )\n    params[""params""][""lam""].update(\n        PyTreeCheckpointer().restore(lam)[""model""][""params""][""params""]\n    )\n    return params\n",python,tab
+110,379547,"genie.py",137,0,"",python,selection_mouse
+111,380221,"genie.py",136,0,"",python,selection_mouse
+112,380223,"genie.py",135,0,"",python,selection_command
+113,416823,"genie.py",3717,0,"",python,selection_mouse
+114,421437,"genie.py",0,0,"",python,tab
+115,421439,"genie.py",4609,0,"",python,selection_mouse
+116,421439,"genie.py",4608,0,"",python,selection_command
+117,422058,"genie.py",4610,0,"",python,selection_mouse
+118,422815,"genie.py",4609,0,"",python,selection_mouse
+119,422859,"genie.py",4608,0,"",python,selection_command
+120,424565,"genie.py",4609,0,"\n        # --- Run MaskGIT loop ---",python,content
+121,424582,"genie.py",4618,0,"",python,selection_command
+122,426041,"genie.py",4608,0,"",python,selection_command
+123,430728,"genie.py",4627,0,"",python,selection_mouse
+124,431729,"genie.py",4609,0,"",python,selection_mouse
+125,431731,"genie.py",4608,0,"",python,selection_command
+126,434063,"genie.py",4862,0,"",python,selection_mouse
+127,434065,"genie.py",4861,0,"",python,selection_command
+128,450431,"genie.py",8719,0,"",python,selection_mouse
+129,450435,"genie.py",8718,0,"",python,selection_command
+130,453692,"genie.py",8717,0,"",python,selection_command
+131,453865,"genie.py",8716,0,"",python,selection_command
+132,454067,"genie.py",8715,0,"",python,selection_command
+133,454889,"genie.py",8715,4,"",python,content
+134,459081,"genie.py",4605,0,"",python,selection_mouse
+135,459613,"genie.py",4605,4,"",python,content
+136,463564,"genie.py",2739,0,"",python,selection_mouse
+137,463992,"genie.py",2739,4,"",python,content
+138,510207,"genie.py",2771,0,"",python,selection_mouse
+139,510223,"genie.py",2770,0,"",python,selection_command
+140,518711,"genie.py",5103,0,"",python,selection_mouse
+141,519650,"genie.py",5102,0,"",python,selection_command
+142,527244,"genie.py",5102,1,"=",python,content
+143,529180,"genie.py",5024,0,"",python,selection_mouse
+144,529413,"genie.py",5024,1," ",python,selection_mouse
+145,529413,"genie.py",5024,2," a",python,selection_mouse
+146,529414,"genie.py",5024,3," an",python,selection_mouse
+147,529428,"genie.py",5024,4," and",python,selection_mouse
+148,529441,"genie.py",5024,5," and ",python,selection_mouse
+149,529459,"genie.py",5024,7," and fu",python,selection_mouse
+150,529504,"genie.py",5024,8," and fut",python,selection_mouse
+151,529545,"genie.py",5024,10," and futur",python,selection_mouse
+152,529546,"genie.py",5024,11," and future",python,selection_mouse
+153,529565,"genie.py",5024,12," and future ",python,selection_mouse
+154,529591,"genie.py",5024,13," and future f",python,selection_mouse
+155,529639,"genie.py",5024,14," and future fr",python,selection_mouse
+156,529648,"genie.py",5024,15," and future fra",python,selection_mouse
+157,529711,"genie.py",5024,16," and future fram",python,selection_mouse
+158,530331,"genie.py",5024,15," and future fra",python,selection_mouse
+159,530372,"genie.py",5024,14," and future fr",python,selection_mouse
+160,530418,"genie.py",5024,13," and future f",python,selection_mouse
+161,530626,"genie.py",5024,12," and future ",python,selection_mouse
+162,530676,"genie.py",5024,11," and future",python,selection_mouse
+163,530766,"genie.py",5024,10," and futur",python,selection_mouse
+164,532232,"genie.py",5053,0,"=",python,content
+165,532232,"genie.py",5052,1,"",python,content
+166,533650,"genie.py",5024,76," and future frames (i.e., t == step_t)\n            mask = jnp.arange(seq_len",python,selection_command
+167,534146,"genie.py",5024,11," and future",python,selection_command
+168,534980,"genie.py",5034,0,"",python,selection_command
+169,535115,"genie.py",5033,0,"",python,selection_command
+170,535620,"genie.py",5032,0,"",python,selection_command
+171,535635,"genie.py",5031,0,"",python,selection_command
+172,535671,"genie.py",5030,0,"",python,selection_command
+173,535717,"genie.py",5029,0,"",python,selection_command
+174,535759,"genie.py",5028,0,"",python,selection_command
+175,535759,"genie.py",5027,0,"",python,selection_command
+176,535826,"genie.py",5026,0,"",python,selection_command
+177,535947,"genie.py",5025,0,"",python,selection_command
+178,536348,"genie.py",5025,4,"",python,content
+179,536699,"genie.py",5025,7,"",python,content
+180,537512,"genie.py",5026,0,"",python,selection_command
+181,538034,"genie.py",5027,0,"",python,selection_command
+182,538059,"genie.py",5028,0,"",python,selection_command
+183,538081,"genie.py",5029,0,"",python,selection_command
+184,538104,"genie.py",5030,0,"",python,selection_command
+185,538255,"genie.py",5031,0,"",python,selection_command
+186,538696,"genie.py",5030,0,"",python,selection_command
+187,539040,"genie.py",5030,1,"",python,content
+188,554120,"genie.py",5030,1," ",python,selection_mouse
+189,554121,"genie.py",5030,3," (i",python,selection_mouse
+190,554144,"genie.py",5030,5," (i.e",python,selection_mouse
+191,554160,"genie.py",5030,7," (i.e.,",python,selection_mouse
+192,554194,"genie.py",5030,9," (i.e., t",python,selection_mouse
+193,554241,"genie.py",5030,65," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == st",python,selection_mouse
+194,554286,"genie.py",5030,66," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == ste",python,selection_mouse
+195,554286,"genie.py",5030,67," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step",python,selection_mouse
+196,554291,"genie.py",5030,68," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_",python,selection_mouse
+197,554326,"genie.py",5030,69," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t",python,selection_mouse
+198,554372,"genie.py",5030,70," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t ",python,selection_mouse
+199,554415,"genie.py",5030,71," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t #",python,selection_mouse
+200,554416,"genie.py",5030,72," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t # ",python,selection_mouse
+201,554462,"genie.py",5030,73," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t # (",python,selection_mouse
+202,554502,"genie.py",5030,74," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t # (S",python,selection_mouse
+203,554548,"genie.py",5030,75," (i.e., t == step_t)\n            mask = jnp.arange(seq_len) == step_t # (S,",python,selection_mouse
+204,554589,"genie.py",5030,20," (i.e., t == step_t)",python,selection_mouse
+205,554893,"genie.py",5050,0,"",python,selection_mouse
+206,554895,"genie.py",5049,0,"",python,selection_command
+207,555540,"genie.py",5050,0,"",python,selection_mouse
+208,555547,"genie.py",5049,0,"",python,selection_command
+209,555682,"genie.py",5049,1,")",python,selection_mouse
+210,555691,"genie.py",5050,0,"",python,selection_command
+211,555711,"genie.py",5047,3,"_t)",python,selection_mouse
+212,555727,"genie.py",5046,4,"p_t)",python,selection_mouse
+213,555742,"genie.py",5045,5,"ep_t)",python,selection_mouse
+214,555772,"genie.py",5044,6,"tep_t)",python,selection_mouse
+215,555819,"genie.py",5042,8," step_t)",python,selection_mouse
+216,555864,"genie.py",5041,9,"= step_t)",python,selection_mouse
+217,555865,"genie.py",5040,10,"== step_t)",python,selection_mouse
+218,555865,"genie.py",5039,11," == step_t)",python,selection_mouse
+219,555866,"genie.py",5038,12,"t == step_t)",python,selection_mouse
+220,555866,"genie.py",5037,13," t == step_t)",python,selection_mouse
+221,555909,"genie.py",5036,14,", t == step_t)",python,selection_mouse
+222,555955,"genie.py",5035,15,"., t == step_t)",python,selection_mouse
+223,555957,"genie.py",5034,16,"e., t == step_t)",python,selection_mouse
+224,556008,"genie.py",5033,17,".e., t == step_t)",python,selection_mouse
+225,556060,"genie.py",5032,18,"i.e., t == step_t)",python,selection_mouse
+226,556205,"genie.py",5031,19,"(i.e., t == step_t)",python,selection_mouse
+227,557425,"genie.py",5031,19,"",python,content
+228,557442,"genie.py",5030,0,"",python,selection_command
+229,558985,"genie.py",5030,1,"",python,content
+230,558990,"genie.py",5029,0,"",python,selection_command
+231,565946,"genie.py",5030,0,"",python,selection_command
+232,566207,"genie.py",5030,0," ",python,content
+233,566209,"genie.py",5031,0,"",python,selection_keyboard
+234,566494,"genie.py",5031,0,"()",python,content
+235,566495,"genie.py",5032,0,"",python,selection_keyboard
+236,566865,"genie.py",5032,0,"f",python,content
+237,566866,"genie.py",5033,0,"",python,selection_keyboard
+238,567012,"genie.py",5033,0,"i",python,content
+239,567013,"genie.py",5034,0,"",python,selection_keyboard
+240,567639,"genie.py",5033,1,"",python,content
+241,567918,"genie.py",5033,0,"u",python,content
+242,567919,"genie.py",5034,0,"",python,selection_keyboard
+243,568068,"genie.py",5034,0,"t",python,content
+244,568069,"genie.py",5035,0,"",python,selection_keyboard
+245,568538,"genie.py",5035,0,"u",python,content
+246,568539,"genie.py",5036,0,"",python,selection_keyboard
+247,569644,"genie.py",5036,0,"r",python,content
+248,569645,"genie.py",5037,0,"",python,selection_keyboard
+249,570091,"genie.py",5037,0,"e",python,content
+250,570092,"genie.py",5038,0,"",python,selection_keyboard
+251,570244,"genie.py",5038,0," ",python,content
+252,570245,"genie.py",5039,0,"",python,selection_keyboard
+253,570387,"genie.py",5039,0,"f",python,content
+254,570388,"genie.py",5040,0,"",python,selection_keyboard
+255,570870,"genie.py",5040,0,"r",python,content
+256,570871,"genie.py",5041,0,"",python,selection_keyboard
+257,571044,"genie.py",5041,0,"a",python,content
+258,571045,"genie.py",5042,0,"",python,selection_keyboard
+259,571167,"genie.py",5042,0,"m",python,content
+260,571168,"genie.py",5043,0,"",python,selection_keyboard
+261,571262,"genie.py",5043,0,"e",python,content
+262,571263,"genie.py",5044,0,"",python,selection_keyboard
+263,571419,"genie.py",5044,0,"s",python,content
+264,571419,"genie.py",5045,0,"",python,selection_keyboard
+265,571497,"genie.py",5045,0," ",python,content
+266,571498,"genie.py",5046,0,"",python,selection_keyboard
+267,571621,"genie.py",5046,0,"a",python,content
+268,571621,"genie.py",5047,0,"",python,selection_keyboard
+269,572234,"genie.py",5047,0,"r",python,content
+270,572235,"genie.py",5048,0,"",python,selection_keyboard
+271,572407,"genie.py",5048,0,"e",python,content
+272,572408,"genie.py",5049,0,"",python,selection_keyboard
+273,572521,"genie.py",5049,0," ",python,content
+274,572522,"genie.py",5050,0,"",python,selection_keyboard
+275,572652,"genie.py",5050,0,"m",python,content
+276,572653,"genie.py",5051,0,"",python,selection_keyboard
+277,572711,"genie.py",5051,0,"a",python,content
+278,572711,"genie.py",5052,0,"",python,selection_keyboard
+279,572755,"genie.py",5052,0,"s",python,content
+280,572755,"genie.py",5053,0,"",python,selection_keyboard
+281,572973,"genie.py",5053,0,"k",python,content
+282,572974,"genie.py",5054,0,"",python,selection_keyboard
+283,573302,"genie.py",5054,0,"e",python,content
+284,573303,"genie.py",5055,0,"",python,selection_keyboard
+285,573757,"genie.py",5055,0,"d",python,content
+286,573758,"genie.py",5056,0,"",python,selection_keyboard
+287,574057,"genie.py",5056,0," ",python,content
+288,574058,"genie.py",5057,0,"",python,selection_keyboard
+289,583490,"genie.py",5057,0,"by default",python,content
+290,583880,"genie.py",5067,0," ",python,content
+291,583882,"genie.py",5068,0,"",python,selection_keyboard
+292,584561,"genie.py",5068,0,"u",python,content
+293,584562,"genie.py",5069,0,"",python,selection_keyboard
+294,584664,"genie.py",5069,0,"s",python,content
+295,584665,"genie.py",5070,0,"",python,selection_keyboard
+296,584757,"genie.py",5070,0,"i",python,content
+297,584758,"genie.py",5071,0,"",python,selection_keyboard
+298,584844,"genie.py",5071,0,"n",python,content
+299,584844,"genie.py",5072,0,"",python,selection_keyboard
+300,584936,"genie.py",5072,0,"g",python,content
+301,584937,"genie.py",5073,0,"",python,selection_keyboard
+302,585028,"genie.py",5073,0," ",python,content
+303,585029,"genie.py",5074,0,"",python,selection_keyboard
+304,585224,"genie.py",5074,0,"c",python,content
+305,585225,"genie.py",5075,0,"",python,selection_keyboard
+306,585771,"genie.py",5075,0,"a",python,content
+307,585772,"genie.py",5076,0,"",python,selection_keyboard
+308,585922,"genie.py",5076,0,"u",python,content
+309,585923,"genie.py",5077,0,"",python,selection_keyboard
+310,586114,"genie.py",5077,0,"s",python,content
+311,586115,"genie.py",5078,0,"",python,selection_keyboard
+312,586452,"genie.py",5078,0,"a",python,content
+313,586453,"genie.py",5079,0,"",python,selection_keyboard
+314,586586,"genie.py",5079,0,"l",python,content
+315,586587,"genie.py",5080,0,"",python,selection_keyboard
+316,586664,"genie.py",5080,0," ",python,content
+317,586665,"genie.py",5081,0,"",python,selection_keyboard
+318,586772,"genie.py",5081,0,"m",python,content
+319,586773,"genie.py",5082,0,"",python,selection_keyboard
+320,586909,"genie.py",5082,0,"a",python,content
+321,586910,"genie.py",5083,0,"",python,selection_keyboard
+322,586975,"genie.py",5083,0,"s",python,content
+323,586976,"genie.py",5084,0,"",python,selection_keyboard
+324,587079,"genie.py",5084,0,"k",python,content
+325,587080,"genie.py",5085,0,"",python,selection_keyboard
+326,587209,"genie.py",5085,0," ",python,content
+327,587210,"genie.py",5086,0,"",python,selection_keyboard
+328,587298,"genie.py",5086,0,"i",python,content
+329,587299,"genie.py",5087,0,"",python,selection_keyboard
+330,587416,"genie.py",5087,0,"n",python,content
+331,587417,"genie.py",5088,0,"",python,selection_keyboard
+332,587462,"genie.py",5088,0," ",python,content
+333,587462,"genie.py",5089,0,"",python,selection_keyboard
+334,589426,"genie.py",5089,0,"S",python,content
+335,589427,"genie.py",5090,0,"",python,selection_keyboard
+336,589557,"genie.py",5090,0,"T",python,content
+337,589558,"genie.py",5091,0,"",python,selection_keyboard
+338,590518,"genie.py",5091,0,"-transformer",python,content
+339,612811,"genie.py",8386,0,"",python,selection_mouse
+340,613438,"genie.py",8378,0,"",python,selection_mouse
+341,620653,"genie.py",8422,0,"",python,selection_mouse
+342,623924,"genie.py",8423,0,"",python,selection_command
+343,625079,"genie.py",8422,1,"",python,content
+344,625303,"genie.py",8422,0,"<",python,content
+345,625304,"genie.py",8423,0,"",python,selection_keyboard
+346,625716,"genie.py",8423,0,"=",python,content
+347,625717,"genie.py",8424,0,"",python,selection_keyboard
+348,627841,"genie.py",8424,0," ",python,content
+349,627843,"genie.py",8425,0,"",python,selection_keyboard
+350,628247,"genie.py",8425,0,"N",python,content
+351,628248,"genie.py",8426,0,"",python,selection_keyboard
+352,628587,"genie.py",8426,0," ",python,content
+353,628588,"genie.py",8427,0,"",python,selection_keyboard
+354,628744,"genie.py",8427,0,"-",python,content
+355,628745,"genie.py",8428,0,"",python,selection_keyboard
+356,637777,"genie.py",8388,0,"",python,selection_mouse
+357,638604,"genie.py",8387,0,"",python,selection_command
+358,639202,"genie.py",8448,0,"\n        ",python,content
+359,639640,"genie.py",8457,0,"f",python,content
+360,639641,"genie.py",8458,0,"",python,selection_keyboard
+361,639733,"genie.py",8458,0,"i",python,content
+362,639734,"genie.py",8459,0,"",python,selection_keyboard
+363,639828,"genie.py",8459,0,"n",python,content
+364,639829,"genie.py",8460,0,"",python,selection_keyboard
+365,639926,"genie.py",8460,0,"a",python,content
+366,639926,"genie.py",8461,0,"",python,selection_keyboard
+367,640014,"genie.py",8461,0,"l",python,content
+368,640016,"genie.py",8462,0,"",python,selection_keyboard
+369,640351,"genie.py",8462,0,"_",python,content
+370,640352,"genie.py",8463,0,"",python,selection_keyboard
+371,642219,"genie.py",8463,0,"t",python,content
+372,642221,"genie.py",8464,0,"",python,selection_keyboard
+373,642352,"genie.py",8464,0,"o",python,content
+374,642353,"genie.py",8465,0,"",python,selection_keyboard
+375,642447,"genie.py",8465,0,"k",python,content
+376,642448,"genie.py",8466,0,"",python,selection_keyboard
+377,643095,"genie.py",8466,0,"e",python,content
+378,643096,"genie.py",8467,0,"",python,selection_keyboard
+379,643188,"genie.py",8467,0,"n",python,content
+380,643191,"genie.py",8468,0,"",python,selection_keyboard
+381,643668,"genie.py",8468,0,"_",python,content
+382,643669,"genie.py",8469,0,"",python,selection_keyboard
+383,643945,"genie.py",8469,0,"p",python,content
+384,643945,"genie.py",8470,0,"",python,selection_keyboard
+385,644099,"genie.py",8470,0,"r",python,content
+386,644100,"genie.py",8471,0,"",python,selection_keyboard
+387,644188,"genie.py",8471,0,"o",python,content
+388,644190,"genie.py",8472,0,"",python,selection_keyboard
+389,644421,"genie.py",8472,0,"b",python,content
+390,644421,"genie.py",8473,0,"",python,selection_keyboard
+391,644658,"genie.py",8473,0,"s",python,content
+392,644659,"genie.py",8474,0,"",python,selection_keyboard
+393,644846,"genie.py",8474,0,"_",python,content
+394,644847,"genie.py",8475,0,"",python,selection_keyboard
+395,645035,"genie.py",8475,0,"f",python,content
+396,645036,"genie.py",8476,0,"",python,selection_keyboard
+397,645165,"genie.py",8476,0,"l",python,content
+398,645166,"genie.py",8477,0,"",python,selection_keyboard
+399,645318,"genie.py",8477,0,"a",python,content
+400,645319,"genie.py",8478,0,"",python,selection_keyboard
+401,645496,"genie.py",8478,0,"t",python,content
+402,645497,"genie.py",8479,0,"",python,selection_keyboard
+403,653294,"genie.py",8479,0," = einops.rearrange(final_token_probs, ""b s n -> b (s n)"")",python,content
+404,664215,"genie.py",8556,0,"",python,selection_mouse
+405,665816,"genie.py",8589,0,"",python,selection_mouse
+406,676623,"genie.py",8589,0,"_flat",python,content
+407,677657,"genie.py",8609,0,"",python,selection_mouse
+408,677806,"genie.py",8605,10,"descending",python,selection_mouse
+409,678021,"genie.py",8605,11,"descending=",python,selection_mouse
+410,678028,"genie.py",8605,15,"descending=True",python,selection_mouse
+411,678495,"genie.py",8605,15,"",python,content
+412,678605,"genie.py",8604,1,"",python,content
+413,678911,"genie.py",8603,1,"",python,content
+414,680061,"genie.py",8604,0,"",python,selection_mouse
+415,682310,"genie.py",8704,0,"",python,selection_mouse
+416,682869,"genie.py",8623,0,"",python,selection_mouse
+417,685402,"genie.py",8564,0,"",python,selection_mouse
+418,686903,"genie.py",8563,0,"",python,selection_command
+419,688625,"genie.py",8562,0,"",python,selection_command
+420,688806,"genie.py",8604,0,"\n        ",python,content
+421,689038,"genie.py",8613,0,"m",python,content
+422,689039,"genie.py",8614,0,"",python,selection_keyboard
+423,689135,"genie.py",8614,0,"a",python,content
+424,689136,"genie.py",8615,0,"",python,selection_keyboard
+425,689180,"genie.py",8615,0,"s",python,content
+426,689182,"genie.py",8616,0,"",python,selection_keyboard
+427,689294,"genie.py",8616,0,"k",python,content
+428,689295,"genie.py",8617,0,"",python,selection_keyboard
+429,689847,"genie.py",8617,0,"_",python,content
+430,689849,"genie.py",8618,0,"",python,selection_keyboard
+431,690458,"genie.py",8618,0,"f",python,content
+432,690459,"genie.py",8619,0,"",python,selection_keyboard
+433,690502,"genie.py",8619,0,"l",python,content
+434,690503,"genie.py",8620,0,"",python,selection_keyboard
+435,690655,"genie.py",8620,0,"a",python,content
+436,690656,"genie.py",8621,0,"",python,selection_keyboard
+437,690901,"genie.py",8621,0,"t",python,content
+438,690903,"genie.py",8622,0,"",python,selection_keyboard
+439,691121,"genie.py",8622,0," ",python,content
+440,691122,"genie.py",8623,0,"",python,selection_keyboard
+441,691566,"genie.py",8623,0,"= einops.rearrange(mask, ""b s n -> b (s n)"")",python,content
+442,705699,"genie.py",8613,0,"",python,selection_command
+443,707905,"genie.py",8605,63,"",python,content
+444,707922,"genie.py",8613,0,"",python,selection_command
+445,708062,"genie.py",8682,0,"\n        mask_flat = einops.rearrange(mask, ""b s n -> b (s n)"")",python,content
+446,708063,"genie.py",8691,0,"",python,selection_command
+447,713388,"genie.py",8799,0,"        new_mask = einops.rearrange(new_mask, ""b (s n) -> b s n"")\n",python,content
+448,713389,"genie.py",8784,0,"_flat",python,content
+449,721894,"genie.py",8762,0,"",python,selection_mouse
+450,723325,"genie.py",8762,0,"_",python,content
+451,723326,"genie.py",8763,0,"",python,selection_keyboard
+452,723699,"genie.py",8763,0,"f",python,content
+453,723700,"genie.py",8764,0,"",python,selection_keyboard
+454,723781,"genie.py",8764,0,"l",python,content
+455,723782,"genie.py",8765,0,"",python,selection_keyboard
+456,723967,"genie.py",8765,0,"a",python,content
+457,723968,"genie.py",8766,0,"",python,selection_keyboard
+458,724104,"genie.py",8766,0,"t",python,content
+459,724105,"genie.py",8767,0,"",python,selection_keyboard
+460,725501,"genie.py",8853,0,"_flat",python,content
+461,725502,"genie.py",8858,0,"",python,selection_command
+462,728250,"genie.py",8880,0,"",python,selection_mouse
+463,729814,"genie.py",8878,0,"",python,selection_mouse
+464,731280,"genie.py",8878,0,",",python,content
+465,731281,"genie.py",8879,0,"",python,selection_keyboard
+466,731433,"genie.py",8879,0," ",python,content
+467,731434,"genie.py",8880,0,"",python,selection_keyboard
+468,731757,"genie.py",8880,0,"n",python,content
+469,731758,"genie.py",8881,0,"",python,selection_keyboard
+470,732730,"genie.py",8881,0,"=N",python,content
+471,733666,"genie.py",8885,0,"",python,selection_mouse
+472,770425,"genie.py",9190,0,"",python,selection_mouse
+473,836690,"TERMINAL",0,0,"\r[?1l>]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0",,terminal_output
+474,838681,"TERMINAL",0,0,"git status",,terminal_command
+475,838691,"TERMINAL",0,0,"]633;E;2025-08-03 16:42:10 git status;d4f80100-2225-40b3-b029-72e571a8ef11]633;COn branch genie-sampling\r\nYour branch is up to date with 'origin/genie-sampling'.\r\n\r\nChanges not staged for commit:\r\n  (use ""git add <file>..."" to update what will be committed)\r\n  (use ""git restore <file>..."" to discard changes in working directory)\r\n\tmodified:   genie.py\r\n\r\nno changes added to commit (use ""git add"" and/or ""git commit -a"")\r\n]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0",,terminal_output
+476,843701,"TERMINAL",0,0,"git add genie.py",,terminal_command
+477,843735,"TERMINAL",0,0,"]633;E;2025-08-03 16:42:15 git add genie.py ;d4f80100-2225-40b3-b029-72e571a8ef11]633;C]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0]633;P;Cwd=/home/hk-project-p0023960/tum_cte0515/Projects/tmp/jafar",,terminal_output
+478,851838,"TERMINAL",0,0,"git status",,terminal_command
+479,858510,"TERMINAL",0,0,"git restore --staged genie.py",,terminal_command
+480,864026,"TERMINAL",0,0,"git diff > diff.diff",,terminal_command
+481,866790,"diff.diff",0,0,"diff --git a/genie.py b/genie.py\nindex 50729ab..dfa352c 100644\n--- a/genie.py\n+++ b/genie.py\n@@ -4,6 +4,7 @@ from orbax.checkpoint import PyTreeCheckpointer\n import jax\n import jax.numpy as jnp\n import flax.linen as nn\n+import einops\n \n from models.dynamics import DynamicsMaskGIT\n from models.lam import LatentActionModel\n@@ -87,25 +88,42 @@ class Genie(nn.Module):\n     def sample(\n         self,\n         batch: Dict[str, Any],\n+        seq_len: int,\n         steps: int = 25,\n-        temperature: int = 1,\n+        temperature: float = 1,\n         sample_argmax: bool = False,\n     ) -> Any:\n+        """"""\n+        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n+\n+        - Input frames are tokenized once.\n+        - Future frames are generated autoregressively in token space.\n+        - All frames are detokenized in a single pass.\n+\n+        Note:\n+        - For interactive or step-wise sampling, detokenization should occur after each action.\n+        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n+        - Temporal causal structure is preserved by \n+            a) reapplying the mask before each decoding step.\n+            b) a temporal causal mask is applied within each ST-transformer block.\n+\n+        Dimension keys:\n+            B: batch size  \n+            T: number of input (conditioning) frames  \n+            N: patches per frame  \n+            S: sequence length  \n+            A: action space  \n+            D: model latent dimension\n+        """"""\n         # --- Encode videos and actions ---\n         tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n-        token_idxs = tokenizer_out[""indices""]\n-        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n+        token_idxs = tokenizer_out[""indices""] # (B, T, N)\n+        B, T, N = token_idxs.shape\n+        pad_shape = (B, seq_len - T, N)\n+        pad = jnp.zeros(pad_shape, dtype=token_idxs.dtype)\n+        token_idxs = jnp.concatenate([token_idxs, pad], axis=1) # (B, S, N)\n         action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n \n-        # --- Initialize MaskGIT ---\n-        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n-        init_carry = (\n-            batch[""rng""],\n-            new_frame_idxs,\n-            init_mask,\n-            token_idxs,\n-            action_tokens,\n-        )\n         MaskGITLoop = nn.scan(\n             MaskGITStep,\n             variable_broadcast=""params"",\n@@ -123,13 +141,45 @@ class Genie(nn.Module):\n             sample_argmax=sample_argmax,\n             steps=steps,\n         )\n-        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n-        new_frame_idxs = final_carry[1]\n-        new_frame_pixels = self.tokenizer.decode(\n-            jnp.expand_dims(new_frame_idxs, 1),\n+\n+        def generation_step_fn(carry, step_t):\n+            rng, current_token_idxs = carry\n+            rng, step_rng = jax.random.split(rng)\n+\n+            # Mask current frame (future frames are masked by default using causal mask in ST-transformer)\n+            mask = jnp.arange(seq_len) == step_t # (S,)\n+            mask = jnp.broadcast_to(mask[None, :, None], (B, seq_len, N)) # (B, S, N)\n+            mask = mask.astype(bool)\n+            masked_token_idxs = current_token_idxs * ~mask\n+\n+            # --- Initialize and run MaskGIT loop ---\n+            init_carry_maskgit = (\n+                step_rng,\n+                masked_token_idxs,\n+                mask,\n+                action_tokens,\n+            )\n+            final_carry_maskgit, _ = loop_fn(init_carry_maskgit, jnp.arange(steps))\n+            updated_token_idxs = final_carry_maskgit[1]\n+            new_carry = (rng, updated_token_idxs)\n+            return new_carry, None\n+\n+        # --- Run the autoregressive generation using scan ---\n+        initial_carry = (batch[""rng""], token_idxs)\n+        timesteps_to_scan = jnp.arange(T, seq_len)\n+        final_carry, _ = jax.lax.scan(\n+            generation_step_fn,\n+            initial_carry,\n+            timesteps_to_scan\n+        )\n+        final_token_idxs = final_carry[1]\n+\n+        # --- Decode all tokens at once at the end ---\n+        final_frames = self.tokenizer.decode(\n+            final_token_idxs,\n             video_hw=batch[""videos""].shape[2:4],\n         )\n-        return new_frame_pixels\n+        return final_frames\n \n     def vq_encode(self, batch, training) -> Dict[str, Any]:\n         # --- Preprocess videos ---\n@@ -181,10 +231,13 @@ class MaskGITStep(nn.Module):\n \n         # --- Update mask ---\n         num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n-        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n-        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n+        idx_mask = jnp.arange(final_token_probs.shape[-1]) <= N - num_unmasked_tokens\n+        final_token_probs_flat = einops.rearrange(final_token_probs, ""b s n -> b (s n)"")\n+        sorted_idxs = jnp.argsort(final_token_probs_flat, axis=-1)\n         mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n-        new_mask = mask_update_fn(mask, sorted_idxs)\n+        mask_flat = einops.rearrange(mask, ""b s n -> b (s n)"")\n+        new_mask_flat = mask_update_fn(mask_flat, sorted_idxs)\n+        new_mask = einops.rearrange(new_mask_flat, ""b (s n) -> b s n"", n=N)\n \n         new_carry = (rng, token_idxs, new_mask, action_tokens)\n         return new_carry, None\n",diff,tab
+482,867593,"diff.diff",752,0,"",diff,selection_mouse
+483,868584,"diff.diff",0,0,"",diff,selection_command
+484,868943,"diff.diff",0,32,"diff --git a/genie.py b/genie.py",diff,selection_command
+485,869351,"diff.diff",0,5545,"diff --git a/genie.py b/genie.py\nindex 50729ab..dfa352c 100644\n--- a/genie.py\n+++ b/genie.py\n@@ -4,6 +4,7 @@ from orbax.checkpoint import PyTreeCheckpointer\n import jax\n import jax.numpy as jnp\n import flax.linen as nn\n+import einops\n \n from models.dynamics import DynamicsMaskGIT\n from models.lam import LatentActionModel\n@@ -87,25 +88,42 @@ class Genie(nn.Module):\n     def sample(\n         self,\n         batch: Dict[str, Any],\n+        seq_len: int,\n         steps: int = 25,\n-        temperature: int = 1,\n+        temperature: float = 1,\n         sample_argmax: bool = False,\n     ) -> Any:\n+        """"""\n+        Autoregressively samples up to `seq_len` future frames, following Figure 8 of the paper.\n+\n+        - Input frames are tokenized once.\n+        - Future frames are generated autoregressively in token space.\n+        - All frames are detokenized in a single pass.\n+\n+        Note:\n+        - For interactive or step-wise sampling, detokenization should occur after each action.\n+        - To maintain consistent tensor shapes across timesteps, all current and future frames are decoded at every step.\n+        - Temporal causal structure is preserved by \n+            a) reapplying the mask before each decoding step.\n+            b) a temporal causal mask is applied within each ST-transformer block.\n+\n+        Dimension keys:\n+            B: batch size  \n+            T: number of input (conditioning) frames  \n+            N: patches per frame  \n+            S: sequence length  \n+            A: action space  \n+            D: model latent dimension\n+        """"""\n         # --- Encode videos and actions ---\n         tokenizer_out = self.tokenizer.vq_encode(batch[""videos""], training=False)\n-        token_idxs = tokenizer_out[""indices""]\n-        new_frame_idxs = jnp.zeros_like(token_idxs)[:, 0]\n+        token_idxs = tokenizer_out[""indices""] # (B, T, N)\n+        B, T, N = token_idxs.shape\n+        pad_shape = (B, seq_len - T, N)\n+        pad = jnp.zeros(pad_shape, dtype=token_idxs.dtype)\n+        token_idxs = jnp.concatenate([token_idxs, pad], axis=1) # (B, S, N)\n         action_tokens = self.lam.vq.get_codes(batch[""latent_actions""])\n \n-        # --- Initialize MaskGIT ---\n-        init_mask = jnp.ones_like(token_idxs, dtype=bool)[:, 0]\n-        init_carry = (\n-            batch[""rng""],\n-            new_frame_idxs,\n-            init_mask,\n-            token_idxs,\n-            action_tokens,\n-        )\n         MaskGITLoop = nn.scan(\n             MaskGITStep,\n             variable_broadcast=""params"",\n@@ -123,13 +141,45 @@ class Genie(nn.Module):\n             sample_argmax=sample_argmax,\n             steps=steps,\n         )\n-        final_carry, _ = loop_fn(init_carry, jnp.arange(steps))\n-        new_frame_idxs = final_carry[1]\n-        new_frame_pixels = self.tokenizer.decode(\n-            jnp.expand_dims(new_frame_idxs, 1),\n+\n+        def generation_step_fn(carry, step_t):\n+            rng, current_token_idxs = carry\n+            rng, step_rng = jax.random.split(rng)\n+\n+            # Mask current frame (future frames are masked by default using causal mask in ST-transformer)\n+            mask = jnp.arange(seq_len) == step_t # (S,)\n+            mask = jnp.broadcast_to(mask[None, :, None], (B, seq_len, N)) # (B, S, N)\n+            mask = mask.astype(bool)\n+            masked_token_idxs = current_token_idxs * ~mask\n+\n+            # --- Initialize and run MaskGIT loop ---\n+            init_carry_maskgit = (\n+                step_rng,\n+                masked_token_idxs,\n+                mask,\n+                action_tokens,\n+            )\n+            final_carry_maskgit, _ = loop_fn(init_carry_maskgit, jnp.arange(steps))\n+            updated_token_idxs = final_carry_maskgit[1]\n+            new_carry = (rng, updated_token_idxs)\n+            return new_carry, None\n+\n+        # --- Run the autoregressive generation using scan ---\n+        initial_carry = (batch[""rng""], token_idxs)\n+        timesteps_to_scan = jnp.arange(T, seq_len)\n+        final_carry, _ = jax.lax.scan(\n+            generation_step_fn,\n+            initial_carry,\n+            timesteps_to_scan\n+        )\n+        final_token_idxs = final_carry[1]\n+\n+        # --- Decode all tokens at once at the end ---\n+        final_frames = self.tokenizer.decode(\n+            final_token_idxs,\n             video_hw=batch[""videos""].shape[2:4],\n         )\n-        return new_frame_pixels\n+        return final_frames\n \n     def vq_encode(self, batch, training) -> Dict[str, Any]:\n         # --- Preprocess videos ---\n@@ -181,10 +231,13 @@ class MaskGITStep(nn.Module):\n \n         # --- Update mask ---\n         num_unmasked_tokens = jnp.round(N * (1.0 - unmasked_ratio)).astype(int)\n-        idx_mask = jnp.arange(final_token_probs.shape[-1]) > num_unmasked_tokens\n-        sorted_idxs = jnp.argsort(final_token_probs, axis=-1, descending=True)\n+        idx_mask = jnp.arange(final_token_probs.shape[-1]) <= N - num_unmasked_tokens\n+        final_token_probs_flat = einops.rearrange(final_token_probs, ""b s n -> b (s n)"")\n+        sorted_idxs = jnp.argsort(final_token_probs_flat, axis=-1)\n         mask_update_fn = jax.vmap(lambda msk, ids: msk.at[ids].set(idx_mask))\n-        new_mask = mask_update_fn(mask, sorted_idxs)\n+        mask_flat = einops.rearrange(mask, ""b s n -> b (s n)"")\n+        new_mask_flat = mask_update_fn(mask_flat, sorted_idxs)\n+        new_mask = einops.rearrange(new_mask_flat, ""b (s n) -> b s n"", n=N)\n \n         new_carry = (rng, token_idxs, new_mask, action_tokens)\n         return new_carry, None\n",diff,selection_command
+486,955527,"genie.py",0,0,"",python,tab
+487,957752,"sample.py",0,0,"",python,tab
+488,979644,"sample.py",2976,0,"",python,selection_mouse
+489,979646,"sample.py",2975,0,"",python,selection_command
+490,979751,"sample.py",2975,1,"d",python,selection_mouse
+491,979757,"sample.py",2976,0,"",python,selection_command
+492,979778,"sample.py",2951,25,"\n    return generated_vid",python,selection_mouse
+493,979794,"sample.py",2931,45,"\n        batch\n    )\n    return generated_vid",python,selection_mouse
+494,979813,"sample.py",2895,81,"d_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+495,979829,"sample.py",2785,191," _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+496,979845,"sample.py",2669,307,"vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+497,979874,"sample.py",2615,361," _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+498,979918,"sample.py",2569,407,"-- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+499,979960,"sample.py",2565,411,"\n# --- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+500,979962,"sample.py",2461,515,"    return module.sample(batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax)\n\n# --- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+501,979966,"sample.py",2423,553,"def _sampling_wrapper(module, batch):\n    return module.sample(batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax)\n\n# --- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,selection_mouse
+502,981058,"sample.py",2423,553,"",python,content
+503,981438,"sample.py",2423,0,"def _sampling_wrapper(module, batch):\n    return module.sample(batch, args.seq_len, args.maskgit_steps, args.temperature, args.sample_argmax)\n\n# --- Define autoregressive sampling loop ---\ndef _autoreg_sample(rng, video_batch, action_batch):\n    vid = video_batch[:, : args.start_frame + 1]\n    sampling_fn = jax.jit(nn.apply(_sampling_wrapper, genie)) \n    rng, _rng = jax.random.split(rng)\n    batch = dict(videos=vid, latent_actions=action_batch, rng=_rng)\n    generated_vid = sampling_fn(\n        params,\n        batch\n    )\n    return generated_vid",python,content
+504,983006,"sample.py",2977,0,"",python,selection_mouse
+505,989842,"sample.py",2931,0,"",python,selection_mouse
+506,990505,"sample.py",2951,0,"",python,selection_mouse
+507,991263,"sample.py",2976,0,"",python,selection_mouse
+508,992649,"genie.py",0,0,"",python,tab
+509,1140179,"TERMINAL",0,0,"git push",,terminal_command
+510,1140230,"TERMINAL",0,0,"]633;E;2025-08-03 16:47:12 git push;d4f80100-2225-40b3-b029-72e571a8ef11]633;C",,terminal_output
+511,1141409,"TERMINAL",0,0,"Enumerating objects: 5, done.\r\nCounting objects:  20% (1/5)\rCounting objects:  40% (2/5)\rCounting objects:  60% (3/5)\rCounting objects:  80% (4/5)\rCounting objects: 100% (5/5)\rCounting objects: 100% (5/5), done.\r\nDelta compression using up to 152 threads\r\nCompressing objects:  33% (1/3)\rCompressing objects:  66% (2/3)\rCompressing objects: 100% (3/3)\rCompressing objects: 100% (3/3), done.\r\nWriting objects:  33% (1/3)\rWriting objects:  66% (2/3)\rWriting objects: 100% (3/3)\rWriting objects: 100% (3/3), 1.95 KiB | 667.00 KiB/s, done.\r\nTotal 3 (delta 2), reused 0 (delta 0), pack-reused 0\r\n",,terminal_output
+512,1141563,"TERMINAL",0,0,"remote: Resolving deltas:   0% (0/2)\rremote: Resolving deltas:  50% (1/2)\rremote: Resolving deltas: 100% (2/2)\rremote: Resolving deltas: 100% (2/2), completed with 2 local objects.\r\n",,terminal_output
+513,1141791,"TERMINAL",0,0,"To github.com:maharajamihir/jafar.git\r\n   ec0d8b2..7fb86d7  genie-sampling -> genie-sampling\r\n]0;tum_cte0515@hkn1991:~/Projects/tmp/jafar]633;D;0",,terminal_output
+514,1216550,"sample.py",0,0,"",python,tab
+515,2056143,"genie.py",0,0,"",python,tab

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-1ae480a2-bc7d-4b65-b456-3679a88992ae1752076515942-2025_07_09-17.56.33.471/source.csv

@@ -0,0 +1,19 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,24812,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"5:56:33 PM [info] Activating crowd-code\n5:56:33 PM [info] Recording started\n5:56:33 PM [info] Initializing git provider using file system watchers...\n5:56:33 PM [info] Git repository found\n5:56:33 PM [info] Git provider initialized successfully\n5:56:33 PM [info] Initial git state: [object Object]\n",Log,tab
+3,25057,"TERMINAL",0,0,"idling",,terminal_command
+4,25174,"TERMINAL",0,0,"8\t ",,terminal_output
+5,26149,"TERMINAL",0,0,"9\t ",,terminal_output
+6,27180,"TERMINAL",0,0,"7:00\t ",,terminal_output
+7,28069,"TERMINAL",0,0,"/bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt",,terminal_command
+8,28084,"TERMINAL",0,0,"]633;E;2025-07-09 17:57:01 /bin/python3 /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/printEnvVariablesToFile.py /hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash/envVars.txt;2757b4f9-70d1-4da5-bad7-c693a73e7031]633;C]0;tum_cte0515@hkn1990:/hkfs/home/project/hk-project-p0023960/tum_cte0515/.cursor-server/extensions/ms-python.python-2024.12.3-linux-x64/python_files/deactivate/bash]633;D;0",,terminal_output
+9,28308,"TERMINAL",0,0,"1\t ",,terminal_output
+10,29257,"TERMINAL",0,0,"2\t ",,terminal_output
+11,30300,"TERMINAL",0,0,"3\t ",,terminal_output
+12,31342,"TERMINAL",0,0,"4\t ",,terminal_output
+13,32408,"TERMINAL",0,0,"5\t ",,terminal_output
+14,33443,"TERMINAL",0,0,"Every 1.0s: sinfo_t_idlehkn1990.localdomain: Wed Jul  9 17:57:06 2025Partition dev_cpuonly:     10 nodes idle\rPartition cpuonly:\t 8 nodes idle\rPartition dev_accelerated:\t 0 nodes idle\rPartition accelerated:\t 2 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 7 nodes idle",,terminal_output
+15,33641,"TERMINAL",0,0,"Every 1.0s: sinfo_t_idlehkn1990.localdomain: Wed Jul  9 17:57:06 2025Partition dev_cpuonly:     10 nodes idle\rPartition cpuonly:\t 8 nodes idle\rPartition dev_accelerated:\t 0 nodes idle\rPartition accelerated:\t 2 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 7 nodes idleEvery 1.0s: sinfo_t_idlehkn1990.localdomain: Wed Jul  9 17:57:06 2025Partition dev_cpuonly:     10 nodes idle\rPartition cpuonly:\t 8 nodes idle\rPartition dev_accelerated:\t 0 nodes idle\rPartition accelerated:\t 2 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 7 nodes idle",,terminal_output
+16,33773,"TERMINAL",0,0,"Every 1.0s: sinfo_t_idlehkn1990.localdomain: Wed Jul  9 17:57:07 2025Partition dev_cpuonly:     10 nodes idle\rPartition cpuonly:\t 8 nodes idle\rPartition dev_accelerated:\t 0 nodes idle\rPartition accelerated:\t 2 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 7 nodes idle",,terminal_output
+17,34764,"TERMINAL",0,0,"Every 1.0s: sinfo_t_idlehkn1990.localdomain: Wed Jul  9 17:57:08 2025Partition dev_cpuonly:     10 nodes idle\rPartition cpuonly:\t 8 nodes idle\rPartition dev_accelerated:\t 0 nodes idle\rPartition accelerated:\t 2 nodes idle\rPartition dev_accelerated-h100    :\t 1 nodes idle\rPartition accelerated-h100:\t 0 nodes idle\rPartition large:\t 7 nodes idle",,terminal_output
+18,35879,"TERMINAL",0,0,"9",,terminal_output
+19,36786,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1990:~/Projects/jafar]633;D;0",,terminal_output

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-311b67df-2f46-4805-a16c-07606ec1fc001751032376516-2025_06_27-15.53.06.902/source.csv

@@ -0,0 +1,2 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+2,275,"tasks",0,0,"",Log,tab

927a8af5474e5654810c00ce2e09fd2de87d3e5722f33fa1090d867db114e403/crowd-code-fbdad54f-458b-4a49-b398-aff5c8bb59901759170581539-2025_09_29-20.30.43.940/source.csv

@@ -0,0 +1,155 @@
+Sequence,Time,File,RangeOffset,RangeLength,Text,Language,Type
+1,5,"slurm/jobs/mihir/horeka/jafar_og_reproduction/og_coinrun_dynamics_reproduction.sbatch",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=2-00:00:00\n#SBATCH --cpus-per-task=8\n#SBATCH --gres=gpu:1\n#SBATCH --partition=accelerated\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/jafar_og_reproduction/dynamics/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/jafar_og_reproduction/dynamics/%x_%j.log\n#SBATCH --job-name=train_dynamics_coinrun_og_reproduction\n\n# Log the sbatch script\ncat $0\n\nsource .venv/bin/activate\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\ntags=""coinrun_og dynanmics 10m_dataset repro_mihir""\n\nnpy_records_dir=""/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_coinrun/coinrun_episodes_10m""\n\n# TODO mihir: update the tokenizer and lam checkpoints\ntokenizer_ckpt_dir=""""\nlam_ckpt_dir=""""\n\nCHECKPOINT_DIR=""/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/mihir/jafar_og_reproduction/dynamics/${job_name}/${slurm_job_id}""\nmkdir -p $CHECKPOINT_DIR\n\nenv | grep SLURM\n\nsrun python jasmine/train_dynamics.py \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --tokenizer_checkpoint=""${tokenizer_ckpt_dir}"" \\n    --lam_checkpoint=""${lam_ckpt_dir}"" \\n    --log_image_interval=1000 \\n    --log \\n    --name=""${job_name}_${slurm_job_id}"" \\n    --tags ${tags} \\n    --entity instant-uv \\n    --project jafar \\n    --data_dir $npy_records_dir \\n    --wandb_id $slurm_job_id\n",shellscript,tab
+2,613,"extension-output-pdoom-org.crowd-code-#1-crowd-code",0,0,"8:30:43 PM [info] Activating crowd-code\n8:30:43 PM [info] Recording started\n8:30:43 PM [info] Initializing git provider using file system watchers...\n",Log,tab
+3,1003,"extension-output-pdoom-org.crowd-code-#1-crowd-code",150,0,"8:30:44 PM [info] Git repository found\n8:30:44 PM [info] Git provider initialized successfully\n8:30:44 PM [info] Initial git state: [object Object]\n",Log,content
+4,1594,"slurm/jobs/mihir/horeka/jafar_og_reproduction/og_coinrun_dynamics_reproduction.sbatch",0,0,"",shellscript,tab
+5,13198,"slurm/jobs/mihir/horeka/minecraft/default_runs/train_tokenizer_8_nodes.sbatch",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=8\n#SBATCH --ntasks-per-node=4\n#SBATCH --time=48:00:00\n#SBATCH --partition=accelerated\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:4\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/%x_%j.log\n#SBATCH --job-name=train_tokenizer_batch_size_scaling_8_node\n#SBATCH --mem=100G\n#SBATCH --requeue\n#SBATCH --signal=b:usr1@300 # 5 min before timeout\n\n# --- signal trap to requeue job before timeout ---\nrequeue_job() {\n    echo ""[$(date)] caught sigusr1 (timeout warning), requeueing slurm job $SLURM_JOB_ID...""\n    # optional: trigger checkpoint saving here\n    # e.g., touch $checkpoint_dir/requeue_trigger\n    scontrol requeue $SLURM_JOB_ID\n    exit 0\n}\n\ntrap requeue_job sigusr1\n\n# set checkpoint flag based on restart count\nrestart_count=$(scontrol show job $SLURM_JOB_ID | grep -o 'Restarts=[0-9]*' | cut -d'=' -f2)\n\nif [ $restart_count -eq 0 ]; then\n    restore_ckpt_flag=""--no-restore-ckpt""\nelse\n    restore_ckpt_flag=""--restore-ckpt""\nfi\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\narray_records_dir_train=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_minecraft/open_ai_minecraft_arrayrecords_chunked_train_val_split/train\narray_records_dir_val=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_minecraft/open_ai_minecraft_arrayrecords_chunked_train_val_split/val\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/mihir/$job_name/$slurm_job_id\nmkdir -p $CHECKPOINT_DIR\n\nenv | grep SLURM\n\nsrun python jasmine/train_tokenizer.py \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=384 \\n    --max_lr=3e-4 \\n    --log_image_interval=1000 \\n    --log_checkpoint_interval=1000 \\n    --log \\n    --patch_size=16 \\n    --name=minecraft-tokenizer-default-$slurm_job_id \\n    --tags minecraft-tokenizer default \\n    --entity instant-uv \\n    --project jafar \\n    --data_dir $array_records_dir_train \\n    --val_data_dir $array_records_dir_val &\n\nchild_pid=$!\n\nwait $child_pid\n",shellscript,tab
+6,30897,"TERMINAL",0,0,"scancel 3528956",,terminal_command
+7,30938,"TERMINAL",0,0,"]633;C]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+8,32401,"TERMINAL",0,0,"queue",,terminal_command
+9,32502,"TERMINAL",0,0,"]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1993.localdomain: Mon Sep 29 20:31:16 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3528956 accelerat train_la tum_cte0 CG    7:26:35\t  1 hkn08073528968 accelerat train_la tum_cte0 PD\t0:00\t  8 (Priority)3528969 accelerat train_to tum_cte0 PD\t0:00\t  8 (Priority)3528955 accelerat train_to tum_cte0  R    7:26:37\t  1 hkn0709",,terminal_output
+10,33557,"TERMINAL",0,0,"78",,terminal_output
+11,34573,"TERMINAL",0,0,"89",,terminal_output
+12,35644,"TERMINAL",0,0,"940",,terminal_output
+13,36671,"TERMINAL",0,0,"201",,terminal_output
+14,37746,"TERMINAL",0,0,"12",,terminal_output
+15,38758,"TERMINAL",0,0,"23",,terminal_output
+16,39787,"TERMINAL",0,0,"34",,terminal_output
+17,40834,"TERMINAL",0,0,"45",,terminal_output
+18,41911,"TERMINAL",0,0,"56",,terminal_output
+19,42943,"TERMINAL",0,0,"67",,terminal_output
+20,43627,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+21,48614,"TERMINAL",0,0,"source .venv/bin/activate",,terminal_command
+22,54309,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=48:00:00\n#SBATCH --partition=accelerated\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:1\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/coinrun/lam/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/coinrun/lam/%x_%j.log\n#SBATCH --job-name=train_lam_default\n#SBATCH --requeue\n#SBATCH --signal=b:usr1@300 # 5 min before timeout\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\narray_records_dir_train=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_coinrun/coinrun_episodes_10m_gt_actions_split/train\narray_records_dir_val=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_coinrun/coinrun_episodes_10m_gt_actions_split/val\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/big-runs/lam/$job_name/$slurm_job_id\nmkdir -p $CHECKPOINT_DIR\n\nenv | grep SLURM\n\nsrun python jasmine/train_lam.py \\n    --save_ckpt \\n    --image_height=64 \\n    --image_width=64 \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=120 \\n    --log \\n    --name=coinrun-lam-default-$slurm_job_id \\n    --tags lam coinrun default \\n    --entity instant-uv \\n    --project jafar \\n    --data_dir $array_records_dir_train \\n    --val_data_dir $array_records_dir_val \n",shellscript,tab
+23,56801,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1228,0,"",shellscript,selection_mouse
+24,57525,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1272,0,"",shellscript,selection_mouse
+25,57528,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1271,0,"",shellscript,selection_command
+26,58062,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1260,0,"",shellscript,selection_mouse
+27,58076,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1259,0,"",shellscript,selection_command
+28,58622,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1260,0,"\n    ",shellscript,content
+29,59305,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1265,0,"-",shellscript,content
+30,59307,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1266,0,"",shellscript,selection_keyboard
+31,59433,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1266,0,"-",shellscript,content
+32,59435,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1267,0,"",shellscript,selection_keyboard
+33,59632,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1267,0,"m",shellscript,content
+34,59634,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1268,0,"",shellscript,selection_keyboard
+35,59716,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1268,0,"a",shellscript,content
+36,59717,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1269,0,"",shellscript,selection_keyboard
+37,59899,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1269,0,"x",shellscript,content
+38,59900,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1270,0,"",shellscript,selection_keyboard
+39,60820,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1270,0,"_",shellscript,content
+40,60822,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1271,0,"",shellscript,selection_keyboard
+41,61106,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1271,0,"l",shellscript,content
+42,61108,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1272,0,"",shellscript,selection_keyboard
+43,61293,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1272,0,"r",shellscript,content
+44,61295,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1273,0,"",shellscript,selection_keyboard
+45,62052,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1273,0,"=",shellscript,content
+46,62054,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1274,0,"",shellscript,selection_keyboard
+47,63013,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1274,0,"1",shellscript,content
+48,63017,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1275,0,"",shellscript,selection_keyboard
+49,63419,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1274,1,"",shellscript,content
+50,63628,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1274,0,"3",shellscript,content
+51,63629,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1275,0,"",shellscript,selection_keyboard
+52,65582,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1275,0,"e",shellscript,content
+53,65585,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1276,0,"",shellscript,selection_keyboard
+54,65720,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1276,0,"-",shellscript,content
+55,65721,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1277,0,"",shellscript,selection_keyboard
+56,66012,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1277,0,"5",shellscript,content
+57,66013,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1278,0,"",shellscript,selection_keyboard
+58,67875,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1278,0," ",shellscript,content
+59,67877,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1279,0,"",shellscript,selection_keyboard
+60,68053,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1279,0,"\",shellscript,content
+61,68054,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1280,0,"",shellscript,selection_keyboard
+62,68533,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1279,0,"",shellscript,selection_command
+63,69773,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1275,0,"",shellscript,selection_mouse
+64,75236,"jasmine/train_lam.py",0,0,"import os\n\nos.environ.setdefault(""XLA_PYTHON_CLIENT_MEM_FRACTION"", ""0.98"")\n\nfrom dataclasses import dataclass, field\nimport itertools\nfrom typing import cast, Optional\n\nimport einops\nfrom jax.sharding import Mesh, PartitionSpec, NamedSharding\nfrom jax.experimental.mesh_utils import create_device_mesh\nimport optax\nimport orbax.checkpoint as ocp\nimport numpy as np\nimport dm_pix as pix\nimport jax\nimport jax.numpy as jnp\nimport tyro\nimport wandb\nimport grain\nimport flax.nnx as nnx\n\nfrom models.lam import LatentActionModel\nfrom utils.dataloader import get_dataloader\nfrom utils.train_utils import (\n    get_lr_schedule,\n    count_parameters_by_component,\n    print_mem_stats,\n    print_compiled_memory_stats,\n    print_compiled_cost_analysis,\n)\n\n\n@dataclass\nclass Args:\n    # Experiment\n    num_steps: int = 200_000\n    seed: int = 0\n    seq_len: int = 16\n    image_channels: int = 3\n    image_height: int = 90\n    image_width: int = 160\n    data_dir: str = """"\n    save_ckpt: bool = False\n    restore_ckpt: bool = False\n    # Optimization\n    batch_size: int = 36\n    vq_beta: float = 0.25\n    init_lr: float = 0.0\n    max_lr: float = 3e-5\n    decay_end: float = 0.0\n    wsd_decay_steps: int = (\n        10000  # NOTE: wsd_decay_steps will only be used when using a wsd-schedule\n    )\n    warmup_steps: int = 5000\n    lr_schedule: str = ""wsd""  # supported options: wsd, cos\n    vq_reset_thresh: int = 50\n    # LAM\n    model_dim: int = 512\n    ffn_dim: int = 2048\n    latent_dim: int = 32\n    num_latents: int = 6\n    patch_size: int = 16\n    num_blocks: int = 4\n    num_heads: int = 8\n    dropout: float = 0.0\n    codebook_dropout: float = 0.0\n    param_dtype = jnp.float32\n    dtype = jnp.bfloat16\n    # Logging\n    log: bool = False\n    entity: str = """"\n    project: str = """"\n    name: str = ""train_lam""\n    tags: list[str] = field(default_factory=lambda: [""lam""])\n    log_interval: int = 5\n    log_image_interval: int = 250\n    ckpt_dir: str = """"\n    log_checkpoint_interval: int = 10000\n    log_checkpoint_keep_period: int = 20000\n    val_data_dir: str = """"\n    val_interval: int = 20_000\n    val_steps: int = 50\n    wandb_id: str = """"\n    use_flash_attention: bool = True\n\n\ndef build_model(args: Args, rng: jax.Array) -> tuple[LatentActionModel, jax.Array]:\n    rng, _rng = jax.random.split(rng)\n    rngs = nnx.Rngs(_rng)\n    return (\n        LatentActionModel(\n            in_dim=args.image_channels,\n            model_dim=args.model_dim,\n            ffn_dim=args.ffn_dim,\n            latent_dim=args.latent_dim,\n            num_latents=args.num_latents,\n            patch_size=args.patch_size,\n            num_blocks=args.num_blocks,\n            num_heads=args.num_heads,\n            dropout=args.dropout,\n            codebook_dropout=args.codebook_dropout,\n            param_dtype=args.param_dtype,\n            dtype=args.dtype,\n            use_flash_attention=args.use_flash_attention,\n            rngs=rngs,\n        ),\n        rng,\n    )\n\n\ndef build_optimizer(model: LatentActionModel, args: Args) -> nnx.ModelAndOptimizer:\n    lr_schedule = get_lr_schedule(\n        args.lr_schedule,\n        args.init_lr,\n        args.max_lr,\n        args.decay_end,\n        args.num_steps,\n        args.warmup_steps,\n        args.wsd_decay_steps,\n    )\n    tx = optax.adamw(\n        learning_rate=lr_schedule,\n        b1=0.9,\n        b2=0.9,\n        weight_decay=1e-4,\n        mu_dtype=args.param_dtype,  # moments in full precision\n    )\n    optimizer = nnx.ModelAndOptimizer(model, tx)\n    return optimizer\n\n\ndef build_mesh_and_sharding(\n    num_devices: int,\n) -> tuple[Mesh, NamedSharding, NamedSharding]:\n    device_mesh_arr = create_device_mesh((num_devices,))\n    mesh = Mesh(devices=device_mesh_arr, axis_names=(""data"",))\n    replicated_sharding = NamedSharding(mesh, PartitionSpec())\n    videos_sharding = NamedSharding(mesh, PartitionSpec(""data"", None, None, None, None))\n    return mesh, replicated_sharding, videos_sharding\n\n\ndef shard_optimizer_states(\n    optimizer: nnx.ModelAndOptimizer, replicated_sharding: NamedSharding\n) -> None:\n    model_state = nnx.state(optimizer.model)\n    model_sharded_state = jax.lax.with_sharding_constraint(\n        model_state, replicated_sharding\n    )\n    nnx.update(optimizer.model, model_sharded_state)\n    optimizer_state = nnx.state(optimizer, nnx.optimizer.OptState)\n    optimizer_sharded_state = jax.lax.with_sharding_constraint(\n        optimizer_state, replicated_sharding\n    )\n    nnx.update(optimizer, optimizer_sharded_state)\n\n\ndef build_dataloader(args: Args, data_dir: str) -> grain.DataLoaderIterator:\n    image_shape = (args.image_height, args.image_width, args.image_channels)\n    array_record_files = [\n        os.path.join(data_dir, x)\n        for x in os.listdir(data_dir)\n        if x.endswith("".array_record"")\n    ]\n    grain_dataloader = get_dataloader(\n        array_record_files,\n        args.seq_len,\n        # NOTE: We deliberately pass the global batch size\n        # The dataloader shards the dataset across all processes\n        args.batch_size,\n        *image_shape,\n        num_workers=8,\n        prefetch_buffer_size=1,\n        seed=args.seed,\n    )\n    initial_state = grain_dataloader._create_initial_state()\n    grain_iterator = grain.DataLoaderIterator(grain_dataloader, initial_state)\n    return grain_iterator\n\n\ndef build_checkpoint_manager(args: Args) -> Optional[ocp.CheckpointManager]:\n    if args.restore_ckpt or args.save_ckpt:\n        handler_registry = ocp.handlers.DefaultCheckpointHandlerRegistry()\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeSave, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""model_state"", ocp.args.PyTreeRestore, ocp.handlers.PyTreeCheckpointHandler\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointSave,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        handler_registry.add(\n            ""train_dataloader_state"",\n            grain.checkpoint.CheckpointRestore,\n            cast(ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler),\n        )\n        if args.val_data_dir:\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointSave,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n            handler_registry.add(\n                ""val_dataloader_state"",\n                grain.checkpoint.CheckpointRestore,\n                cast(\n                    ocp.handlers.CheckpointHandler, grain.checkpoint.CheckpointHandler\n                ),\n            )\n        checkpoint_options = ocp.CheckpointManagerOptions(\n            save_interval_steps=args.log_checkpoint_interval,\n            max_to_keep=3,\n            keep_period=args.log_checkpoint_keep_period,\n            step_format_fixed_length=6,\n            cleanup_tmp_directories=True,\n        )\n        checkpoint_manager = ocp.CheckpointManager(\n            args.ckpt_dir,\n            options=checkpoint_options,\n            handler_registry=handler_registry,\n        )\n        return checkpoint_manager\n    else:\n        return None\n\n\ndef restore_checkpoint_if_needed(\n    args: Args,\n    checkpoint_manager: Optional[ocp.CheckpointManager],\n    optimizer: nnx.ModelAndOptimizer,\n    train_iterator: grain.DataLoaderIterator,\n    val_iterator: Optional[grain.DataLoaderIterator],\n    restore_step: Optional[int] = None,\n) -> tuple[\n    int, nnx.ModelAndOptimizer, grain.DataLoaderIterator, grain.DataLoaderIterator\n]:\n    step = 0\n    if checkpoint_manager and restore_step is None:\n        restore_step = checkpoint_manager.latest_step()\n    if args.restore_ckpt:\n        assert checkpoint_manager is not None\n        abstract_optimizer = nnx.eval_shape(lambda: optimizer)\n        abstract_optimizer_state = nnx.state(abstract_optimizer)\n        if val_iterator:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n                val_dataloader_state=grain.checkpoint.CheckpointRestore(val_iterator),  # type: ignore\n            )\n        else:\n            restore_args = ocp.args.Composite(\n                model_state=ocp.args.PyTreeRestore(abstract_optimizer_state),  # type: ignore\n                train_dataloader_state=grain.checkpoint.CheckpointRestore(train_iterator),  # type: ignore\n            )\n        restored = checkpoint_manager.restore(restore_step, args=restore_args)\n        restored_optimizer_state = restored[""model_state""]\n        nnx.update(optimizer, restored_optimizer_state)\n        train_iterator = restored[""train_dataloader_state""]\n        if val_iterator:\n            val_iterator = restored[""val_dataloader_state""]\n        step = restore_step or 0\n        print(f""Restored dataloader and model state from step {step}"")\n    return step, optimizer, train_iterator, val_iterator\n\n\ndef enable_sowing(lam: LatentActionModel) -> None:\n    for model in [lam.encoder, lam.decoder]:\n        setattr(model, ""sow_logits"", True)\n        for blk in getattr(model, ""blocks"", []):\n            setattr(blk, ""sow_weights"", True)\n            setattr(blk, ""sow_activations"", True)\n\n\ndef main(args: Args) -> None:\n    jax.distributed.initialize()\n    num_devices = jax.device_count()\n    if num_devices == 0:\n        raise ValueError(""No JAX devices found."")\n    print(f""Running on {num_devices} devices."")\n\n    if args.batch_size % num_devices != 0:\n        raise ValueError(\n            f""Global batch size {args.batch_size} must be divisible by ""\n            f""number of devices {num_devices}.""\n        )\n\n    rng = jax.random.key(args.seed)\n\n    # --- Initialize model ---\n    lam, rng = build_model(args, rng)\n\n    # Count parameters\n    _, params, _ = nnx.split(lam, nnx.Param, ...)\n    param_counts = count_parameters_by_component(params)\n\n    if args.log and jax.process_index() == 0:\n        wandb_init_kwargs = {\n            ""entity"": args.entity,\n            ""project"": args.project,\n            ""name"": args.name,\n            ""tags"": args.tags,\n            ""group"": ""debug"",\n            ""config"": args,\n        }\n\n        if args.wandb_id:\n            wandb_init_kwargs.update(\n                {\n                    ""id"": args.wandb_id,\n                    ""resume"": ""allow"",\n                }\n            )\n        wandb.init(**wandb_init_kwargs)\n\n        wandb.config.update({""model_param_count"": param_counts})\n\n    print(""Parameter counts:"")\n    print(param_counts)\n\n    # --- Initialize optimizer ---\n    optimizer = build_optimizer(lam, args)\n    del lam\n\n    # FIXME: switch to create_hybrid_device_mesh for runs spanning multiple nodes\n    _, replicated_sharding, videos_sharding = build_mesh_and_sharding(num_devices)\n\n    shard_optimizer_states(optimizer, replicated_sharding)\n\n    # --- Initialize checkpoint manager ---\n    checkpoint_manager = build_checkpoint_manager(args)\n\n    # --- Create DataLoaderIterator from dataloader ---\n    train_iterator = build_dataloader(args, args.data_dir)\n    val_iterator = None\n    if args.val_data_dir:\n        val_iterator = build_dataloader(args, args.val_data_dir)\n\n    # --- Restore checkpoint ---\n    step, optimizer, train_iterator, val_iterator = restore_checkpoint_if_needed(\n        args, checkpoint_manager, optimizer, train_iterator, val_iterator\n    )\n\n    # --- Define loss and train step (close over args) ---\n    def lam_loss_fn(\n        model: LatentActionModel, inputs: dict, training: bool = True\n    ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n        gt = jnp.asarray(inputs[""videos""], dtype=jnp.float32) / 255.0\n        inputs[""videos""] = gt.astype(args.dtype)\n        outputs = model(inputs, training)\n        outputs[""recon""] = outputs[""recon""].astype(jnp.float32)\n        gt_future_frames = gt[:, 1:]\n        mse = jnp.square(gt_future_frames - outputs[""recon""]).mean()\n        q_loss = jnp.square(jax.lax.stop_gradient(outputs[""emb""]) - outputs[""z""]).mean()\n        commitment_loss = jnp.square(\n            outputs[""emb""] - jax.lax.stop_gradient(outputs[""z""])\n        ).mean()\n        loss = mse + q_loss + args.vq_beta * commitment_loss\n\n        gt_val = gt_future_frames.clip(0, 1).reshape(-1, *gt_future_frames.shape[2:])\n        recon = outputs[""recon""].clip(0, 1).reshape(-1, *outputs[""recon""].shape[2:])\n        psnr = jnp.asarray(pix.psnr(gt_val, recon)).mean()\n        ssim = jnp.asarray(pix.ssim(gt_val, recon)).mean()\n        count_fn = jax.vmap(lambda i: (outputs[""indices""] == i).sum())\n        index_counts = count_fn(jnp.arange(args.num_latents))\n        metrics = dict(\n            loss=loss,\n            mse=mse,\n            q_loss=q_loss,\n            commitment_loss=commitment_loss,\n            psnr=psnr,\n            ssim=ssim,\n            codebook_usage=(index_counts != 0).mean(),\n        )\n        return loss, (outputs[""recon""], index_counts, metrics)\n\n    @nnx.jit(donate_argnums=0)\n    def train_step(\n        optimizer: nnx.ModelAndOptimizer,\n        inputs: dict,\n        action_last_active: jax.Array,\n        rng: jax.Array,\n    ) -> tuple[jax.Array, jax.Array, jax.Array, dict]:\n        def loss_fn(\n            model: LatentActionModel,\n        ) -> tuple[jax.Array, tuple[jax.Array, jax.Array, dict]]:\n            model.train()\n            return lam_loss_fn(model, inputs, training=True)\n\n        # --- Update model ---\n        (loss, (recon, idx_counts, metrics)), grads = nnx.value_and_grad(\n            loss_fn, has_aux=True\n        )(optimizer.model)\n        optimizer.update(grads)\n\n        # --- Reset inactive latent actions ---\n        codebook = optimizer.model.vq.codebook\n        num_codes = len(codebook)\n        active_codes = idx_counts != 0.0\n        action_last_active = jnp.where(active_codes, 0, action_last_active + 1)\n        p_code = active_codes / active_codes.sum()\n        reset_idxs = jax.random.choice(rng, num_codes, shape=(num_codes,), p=p_code)\n        do_reset = action_last_active >= args.vq_reset_thresh\n        new_codebook = jnp.where(\n            jnp.expand_dims(do_reset, -1), codebook[reset_idxs], codebook.value\n        )\n        optimizer.model.vq.codebook.value = new_codebook\n        action_last_active = jnp.where(do_reset, 0, action_last_active)\n        return loss, recon, action_last_active, metrics\n\n    @nnx.jit\n    def val_step(\n        lam: LatentActionModel, inputs: dict\n    ) -> tuple[jax.Array, jax.Array, dict]:\n        lam.eval()\n        (loss, (recon, _, metrics)) = lam_loss_fn(lam, inputs, training=False)\n        return loss, recon, metrics\n\n    def calculate_validation_metrics(val_dataloader, lam):\n        step = 0\n        loss_per_step = []\n        metrics_per_step = []\n        batch = None\n        recon = None\n        for batch in val_dataloader:\n            loss, recon, metrics = val_step(lam, batch)\n            loss_per_step.append(loss)\n            metrics_per_step.append(metrics)\n            step += 1\n            if step > args.val_steps:\n                break\n\n        if step < args.val_steps:\n            print(\n                f""Warning: Your validation dataset is too small to make val_steps many steps. Made {step} steps, expected {args.val_steps}""\n            )\n\n        val_loss = np.mean(loss_per_step)\n        val_metrics = {\n            f""val_{key}"": np.mean([float(m[key]) for m in metrics_per_step])\n            for key in metrics_per_step[0].keys()\n        }\n        val_metrics[""val_loss""] = val_loss\n        return val_metrics, batch, recon\n\n    # --- TRAIN LOOP ---\n    dataloader_train = (\n        {\n            ""videos"": jax.make_array_from_process_local_data(\n                videos_sharding, elem[""videos""]\n            ),\n        }\n        for elem in train_iterator\n    )\n    dataloader_val = None\n    if val_iterator:\n        dataloader_val = (\n            {\n                ""videos"": jax.make_array_from_process_local_data(\n                    videos_sharding, elem[""videos""]\n                ),\n            }\n            for elem in val_iterator\n        )\n    action_last_active = jnp.zeros(args.num_latents, dtype=jnp.int32)\n    if jax.process_index() == 0:\n        first_batch = next(dataloader_train)\n        compiled = train_step.lower(\n            optimizer, first_batch, action_last_active, rng\n        ).compile()\n        print_compiled_memory_stats(compiled.memory_analysis())\n        print_compiled_cost_analysis(compiled.cost_analysis())\n        # Do not skip the first batch during training\n        dataloader_train = itertools.chain([first_batch], dataloader_train)\n    print(f""Starting training from step {step}..."")\n    first_step = step\n    while step < args.num_steps:\n        for batch in dataloader_train:\n            # --- Train step ---\n            rng, _rng = jax.random.split(rng, 2)\n            loss, recon, action_last_active, metrics = train_step(\n                optimizer, batch, action_last_active, _rng\n            )\n            if step == first_step:\n                print_mem_stats(""After params initialized"")\n            step += 1\n\n            # --- Validation loss ---\n            val_results = {}\n            if dataloader_val and step % args.val_interval == 0:\n                print(""Calculating validation metrics..."")\n                val_metrics, val_gt_batch, val_recon = calculate_validation_metrics(\n                    dataloader_val, optimizer.model\n                )\n                print(f""Step {step}, validation loss: {val_metrics['val_loss']}"")\n                val_results = {\n                    ""metrics"": val_metrics,\n                    ""gt_batch"": val_gt_batch,\n                    ""recon"": val_recon,\n                }\n\n            # --- Logging ---\n            if args.log:\n                if step % args.log_interval == 0 and jax.process_index() == 0:\n                    log_dict = {""loss"": loss, ""step"": step, **metrics}\n                    if val_results:\n                        log_dict.update(val_results[""metrics""])\n                    wandb.log(log_dict)\n                if step % args.log_image_interval == 0:\n                    gt_seq = batch[""videos""][0, 1:].astype(jnp.float32) / 255.0\n                    recon_seq = recon[0].clip(0, 1)\n                    comparison_seq = jnp.concatenate((gt_seq, recon_seq), axis=1)\n                    comparison_seq = einops.rearrange(\n                        comparison_seq * 255, ""t h w c -> h (t w) c""\n                    )\n                    if val_results:\n                        val_results[""gt_seq_val""] = (\n                            val_results[""gt_batch""][""videos""][0, 1:].astype(jnp.float32)\n                            / 255.0\n                        )\n                        val_results[""recon_seq_val""] = val_results[""recon""][0].clip(\n                            0, 1\n                        )\n                        val_results[""val_comparison_seq""] = jnp.concatenate(\n                            (val_results[""gt_seq_val""], val_results[""recon_seq_val""]),\n                            axis=1,\n                        )\n                        val_results[""val_comparison_seq""] = einops.rearrange(\n                            val_results[""val_comparison_seq""] * 255,\n                            ""t h w c -> h (t w) c"",\n                        )\n                    # NOTE: Process-dependent control flow deliberately happens\n                    # after indexing operation since it must not contain code\n                    # sections that lead to cross-accelerator communication.\n                    if jax.process_index() == 0:\n                        log_images = dict(\n                            image=wandb.Image(np.asarray(gt_seq[0])),\n                            recon=wandb.Image(np.asarray(recon_seq[0])),\n                            true_vs_recon=wandb.Image(\n                                np.asarray(comparison_seq.astype(np.uint8))\n                            ),\n                        )\n                        if val_results:\n                            log_images.update(\n                                {\n                                    ""val_image"": wandb.Image(\n                                        np.asarray(val_results[""gt_seq_val""][0])\n                                    ),\n                                    ""val_recon"": wandb.Image(\n                                        np.asarray(val_results[""recon_seq_val""][0])\n                                    ),\n                                    ""val_true_vs_recon"": wandb.Image(\n                                        np.asarray(\n                                            val_results[""val_comparison_seq""].astype(\n                                                np.uint8\n                                            )\n                                        )\n                                    ),\n                                }\n                            )\n                        wandb.log(log_images)\n            # --- Checkpointing ---\n            if args.save_ckpt and step % args.log_checkpoint_interval == 0:\n                assert checkpoint_manager is not None\n                optimizer_state = nnx.state(optimizer)\n                if val_iterator:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                        val_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            val_iterator  # type: ignore\n                        ),\n                    )\n                else:\n                    ckpt_manager_args = ocp.args.Composite(\n                        model_state=ocp.args.PyTreeSave(optimizer_state),  # type: ignore\n                        train_dataloader_state=grain.checkpoint.CheckpointSave(  # type: ignore\n                            train_iterator  # type: ignore\n                        ),\n                    )\n                checkpoint_manager.save(step, args=ckpt_manager_args)\n                print(f""Saved checkpoint at step {step}"")\n            if step >= args.num_steps:\n                break\n\n    if checkpoint_manager:\n        checkpoint_manager.close()\n\n\nif __name__ == ""__main__"":\n    args = tyro.cli(Args)\n    main(args)\n",python,tab
+65,80162,"jasmine/train_lam.py",1124,0,"",python,selection_mouse
+66,80346,"jasmine/train_lam.py",1120,6,"max_lr",python,selection_mouse
+67,86383,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",0,0,"",shellscript,tab
+68,89762,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1277,0,"",shellscript,selection_mouse
+69,90944,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1277,1,"6",shellscript,content
+70,93892,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1276,0,"",shellscript,selection_command
+71,94054,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1275,0,"",shellscript,selection_command
+72,94497,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1274,0,"",shellscript,selection_command
+73,95462,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",1274,1,"5",shellscript,content
+74,100499,"TERMINAL",0,0,"sync-runner",,terminal_command
+75,100562,"TERMINAL",0,0,"]633;Csending incremental file list\r\n",,terminal_output
+76,100949,"TERMINAL",0,0,"pyproject.toml\r\n",,terminal_output
+77,101117,"TERMINAL",0,0,"r",,terminal_output
+78,101190,"TERMINAL",0,0,"u",,terminal_output
+79,101359,"TERMINAL",0,0,"n",,terminal_output
+80,101549,"TERMINAL",0,0,"ne",,terminal_output
+81,101602,"TERMINAL",0,0,"r",,terminal_output
+82,101669,"TERMINAL",0,0,"\r\n",,terminal_output
+83,102832,"TERMINAL",0,0,"jasmine/\r\njasmine/genie.py\r\njasmine/sample.py\r\njasmine/train_dynamics.py\r\njasmine/models/\r\njasmine/models/dynamics.py\r\njasmine/utils/\r\njasmine/utils/dataloader_torch.py\r\nslurm/jobs/mihir/horeka/breakout/noise_schedule_runs/full-prec/train_dyn_single_gpu.sh\r\n",,terminal_output
+84,103054,"TERMINAL",0,0,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh\r\nslurm/jobs/mihir/horeka/preprocessing/\r\nslurm/jobs/mihir/horeka/preprocessing/doom_chunked.sh\r\n",,terminal_output
+85,103131,"TERMINAL",0,0,"\r\nsent 115,079 bytes  received 432 bytes  46,204.40 bytes/sec\r\ntotal size is 28,204,848  speedup is 244.17\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+86,103164,"TERMINAL",0,0,"runner",,terminal_command
+87,106589,"TERMINAL",0,0,"git branch",,terminal_command
+88,109444,"TERMINAL",0,0,"dev",,terminal_command
+89,111011,"TERMINAL",0,0,"git branch",,terminal_command
+90,111060,"TERMINAL",0,0,"]633;C[?1h=\r",,terminal_output
+91,111557,"TERMINAL",0,0,"* ablation/use-pytorch-dataloader\r\n  action-mapper\r\n  add-noise-to-combat-exposure-bias\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n:",,terminal_output
+92,113179,"TERMINAL",0,0,"\r\r:",,terminal_output
+93,113356,"TERMINAL",0,0,"\r\r:",,terminal_output
+94,113567,"TERMINAL",0,0,"\r\r:",,terminal_output
+95,114843,"TERMINAL",0,0,"\r\r:",,terminal_output
+96,115448,"TERMINAL",0,0,"\r[?1l>]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+97,118009,"TERMINAL",0,0,"git status",,terminal_command
+98,118056,"TERMINAL",0,0,"]633;C",,terminal_output
+99,118185,"TERMINAL",0,0,"On branch ablation/use-pytorch-dataloader\r\nLast commands done (2 commands done):\r\n   pick ba37453 feat: generate coinrun dataset with val split\r\n   pick faadd10 feat: implemented validation loss for all three models\r\nNext commands to do (26 remaining commands):\r\n   pick 9a17dbb fix: pass val data path to dataloader\r\n   pick 6e69cdb fix typo in image logging\r\n  (use ""git rebase --edit-todo"" to view and edit)\r\nYou are currently editing a commit while rebasing branch 'gt-actions' on 'c7522f2'.\r\n  (use ""git commit --amend"" to amend the current commit)\r\n  (use ""git rebase --continue"" once you are satisfied with your changes)\r\n\r\nChanges not staged for commit:\r\n  (use ""git add <file>..."" to update what will be committed)\r\n  (use ""git restore <file>..."" to discard changes in working directory)\r\n\tmodified:   jasmine/train_dynamics.py\r\n\tmodified:   pyproject.toml\r\n\r\nUntracked files:\r\n  (use ""git add <file>..."" to include in what will be committed)\r\n\tdata/_vizdoom.ini\r\n\tdata/data/\r\n\tdata/jasmine_data/vizdoom/\r\n\tdata/uv.lock\r\n\tdiff.diff\r\n\tdiff2.diff\r\n\tinput_pipeline/\r\n\tjasmine/utils/dataloader_torch.py\r\n\tkiller.sh\r\n\tkiller_partition.sh\r\n\tlog.log\r\n\toverfit_dir.zip\r\n\trequirements-franz.txt\r\n\tsamples/\r\n\tscripts_cremers/\r\n\tslurm/\r\n\ttest.py\r\n\tutils/\r\n\tuv.lock\r\n\r\nno changes added to commit (use ""git add"" and/or ""git commit -a"")\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+100,127406,"TERMINAL",0,0,"git add jasmine/utils/dataloader_torch.py",,terminal_command
+101,127433,"TERMINAL",0,0,"]633;C]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+102,133324,"TERMINAL",0,0,"git commit -am ""added torch dataloader""",,terminal_command
+103,133374,"TERMINAL",0,0,"]633;C",,terminal_output
+104,135068,"TERMINAL",0,0,"black....................................................................",,terminal_output
+105,137270,"TERMINAL",0,0,"Failed\r\n- hook id: black\r\n- files were modified by this hook\r\n\r\nreformatted jasmine/utils/dataloader_torch.py\r\nreformatted jasmine/train_dynamics.py\r\n\r\nAll done! ✨ 🍰 ✨\r\n2 files reformatted.\r\n\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+106,139016,"TERMINAL",0,0,"git add jasmine/utils/dataloader_torch.py",,terminal_command
+107,139065,"TERMINAL",0,0,"]633;C",,terminal_output
+108,139119,"TERMINAL",0,0,"]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+109,140181,"TERMINAL",0,0,"git commit -am ""added torch dataloader""",,terminal_command
+110,140240,"TERMINAL",0,0,"]633;C",,terminal_output
+111,140758,"TERMINAL",0,0,"black....................................................................",,terminal_output
+112,141060,"TERMINAL",0,0,"Passed\r\n",,terminal_output
+113,141172,"TERMINAL",0,0,"[ablation/use-pytorch-dataloader c486cd9] added torch dataloader\r\n 3 files changed, 43 insertions(+), 68 deletions(-)\r\n create mode 100644 jasmine/utils/dataloader_torch.py\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+114,145776,"TERMINAL",0,0,"git branch",,terminal_command
+115,145880,"TERMINAL",0,0,"]633;C[?1h=\r* ablation/use-pytorch-dataloader\r\n  action-mapper\r\n  add-noise-to-combat-exposure-bias\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n:",,terminal_output
+116,146366,"TERMINAL",0,0,"\r/",,terminal_output
+117,149573,"TERMINAL",0,0,"qq",,terminal_output
+118,150440,"TERMINAL",0,0,"* ablation/use-pytorch-dataloader\r\n  action-mapper\r\n  add-noise-to-combat-exposure-bias\r\n  add-wandb-name-and-tags\r\n  before-nnx\r\n  causal-mem-reduce\r\n  causal-spatiotemporal-kv-cache\r\n  causal-st-transformer\r\n  causal-transformer-dynamics-model\r\n  causal-transformer-nnx-no-kv-cache\r\n  coinrun-gt-actions\r\n  convert-to-jax-array-in-iter\r\n  correct-batched-sampling\r\n  dev\r\n  dont-let-tf-see-gpu\r\n:",,terminal_output
+119,151403,"TERMINAL",0,0,"\r[?1l>\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+120,153667,"TERMINAL",0,0,"git checkout add-noise-to-combat-exposure-bias",,terminal_command
+121,153718,"TERMINAL",0,0,"]633;C",,terminal_output
+122,153931,"TERMINAL",0,0,"Switched to branch 'add-noise-to-combat-exposure-bias'\r\nYour branch is up to date with 'origin/add-noise-to-combat-exposure-bias'.\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine",,terminal_output
+123,155370,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",0,0,"Switched from branch 'ablation/use-pytorch-dataloader' to 'add-noise-to-combat-exposure-bias'",shellscript,git_branch_checkout
+124,156237,"TERMINAL",0,0,"runner",,terminal_command
+125,157938,"TERMINAL",0,0,"sync-runner",,terminal_command
+126,158013,"TERMINAL",0,0,"]633;Csending incremental file list\r\n",,terminal_output
+127,163149,"TERMINAL",0,0,"./\r\npyproject.toml\r\n",,terminal_output
+128,164888,"TERMINAL",0,0,"jasmine/\r\njasmine/genie.py\r\njasmine/sample.py\r\njasmine/train_dynamics.py\r\njasmine/models/\r\njasmine/models/dynamics.py\r\njasmine/utils/\r\n\r\nsent 116,290 bytes  received 350 bytes  15,552.00 bytes/sec\r\ntotal size is 28,210,925  speedup is 241.86\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine_jobs",,terminal_output
+129,168168,"TERMINAL",0,0,"sbatch slurm/jobs/mihir/horeka/coinrun/default_runs/train_lam_default.sh",,terminal_command
+130,168201,"TERMINAL",0,0,"]633;CSubmitted batch job 3529697\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine_jobs",,terminal_output
+131,174845,"jasmine/utils/dataloader.py",0,0,"import jax\nimport numpy as np\nimport grain\nfrom typing import Any\nimport pickle\n\n\nclass EpisodeLengthFilter(grain.transforms.Filter):\n    """"""\n    A Grain Filter that keeps only episodes with sufficient length.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the filter with sequence length requirements.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def filter(self, element: Any) -> bool:\n        """"""\n        Filters episodes based on length.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n\n        Returns:\n            True if the episode has sufficient length, False otherwise.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        current_episode_len = element[""sequence_length""]\n        if current_episode_len < self.seq_len:\n            print(\n                f""Filtering out episode with length {current_episode_len}, which is ""\n                f""shorter than the requested sequence length {self.seq_len}.""\n            )\n            return False\n\n        return True\n\n\nclass ProcessEpisodeAndSlice(grain.transforms.RandomMap):\n    """"""\n    A Grain Transformation that combines parsing, slicing, and normalizing.\n    """"""\n\n    def __init__(self, seq_len: int, image_h: int, image_w: int, image_c: int):\n        """"""Initializes the transformation with processing parameters.""""""\n        self.seq_len = seq_len\n        self.image_h = image_h\n        self.image_w = image_w\n        self.image_c = image_c\n\n    def random_map(self, element: dict, rng: np.random.Generator) -> Any:\n        """"""\n        Processes a single raw episode from the data source.\n\n        Args:\n            element: A dictionary representing one record from the DataSource.\n                     Expected to contain 'raw_video' (bytes) and 'sequence_length' (int)\n            rng: A per-record random number generator provided by the Grain sampler.\n\n        Returns:\n            A processed video sequence as a NumPy array with shape\n            (seq_len, height, width, channels) and dtype float32.\n        """"""\n        assert isinstance(element, bytes)\n        element = pickle.loads(element)\n\n        video_shape = (\n            element[""sequence_length""],\n            self.image_h,\n            self.image_w,\n            self.image_c,\n        )\n        episode_tensor = np.frombuffer(element[""raw_video""], dtype=np.uint8)\n        episode_tensor = episode_tensor.reshape(video_shape)\n\n        current_episode_len = episode_tensor.shape[0]\n        if current_episode_len < self.seq_len:\n            raise ValueError(\n                f""Episode length {current_episode_len} is shorter than ""\n                f""requested sequence length {self.seq_len}. This should ""\n                f""have been filtered out.""\n            )\n\n        max_start_idx = current_episode_len - self.seq_len\n\n        start_idx = rng.integers(0, max_start_idx + 1)\n\n        seq = episode_tensor[start_idx : start_idx + self.seq_len]\n\n        data_dict = {""videos"": seq}\n        if ""actions"" in element.keys():\n            actions_tensor = np.array(element[""actions""])\n            actions = actions_tensor[start_idx : start_idx + self.seq_len]\n            data_dict[""actions""] = actions\n\n        return data_dict\n\n\ndef get_dataloader(\n    array_record_paths: list[str],\n    seq_len: int,\n    global_batch_size: int,\n    image_h: int,\n    image_w: int,\n    image_c: int,\n    num_workers: int = 1,\n    prefetch_buffer_size: int = 1,\n    seed: int = 42,\n):\n    """"""\n    Creates a data loading pipeline using Grain.\n    """"""\n    if not array_record_paths:\n        raise ValueError(""array_record_paths list cannot be empty."")\n\n    num_processes = jax.process_count()\n\n    if global_batch_size % num_processes != 0:\n        raise ValueError(\n            f""Global batch size {global_batch_size} must be divisible by ""\n            f""the number of JAX processes {num_processes} for proper sharding.""\n        )\n    per_process_batch_size = global_batch_size // num_processes\n\n    source = grain.sources.ArrayRecordDataSource(array_record_paths)\n\n    sampler = grain.samplers.IndexSampler(\n        num_records=len(source),\n        shard_options=grain.sharding.ShardByJaxProcess(drop_remainder=True),\n        shuffle=True,\n        num_epochs=None,\n        seed=seed,\n    )\n\n    operations = [\n        EpisodeLengthFilter(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        ProcessEpisodeAndSlice(\n            seq_len=seq_len, image_h=image_h, image_w=image_w, image_c=image_c\n        ),\n        grain.transforms.Batch(batch_size=per_process_batch_size, drop_remainder=True),\n    ]\n\n    read_options = grain.ReadOptions(\n        prefetch_buffer_size=prefetch_buffer_size,\n        num_threads=1,\n    )\n    dataloader = grain.DataLoader(\n        data_source=source,\n        sampler=sampler,\n        operations=operations,\n        worker_count=num_workers,\n        worker_buffer_size=1,\n        read_options=read_options,\n    )\n\n    return dataloader\n",python,tab
+132,196824,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh",0,0,"#!/usr/bin/env bash\n\n#SBATCH --nodes=1\n#SBATCH --ntasks-per-node=1\n#SBATCH --time=48:00:00\n#SBATCH --partition=accelerated\n#SBATCH --cpus-per-task=5\n#SBATCH --gres=gpu:1\n#SBATCH --output=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/coinrun/tokenizer/%x_%j.log\n#SBATCH --error=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/logs/logs_mihir/coinrun/tokenizer/%x_%j.log\n#SBATCH --job-name=train_tokenizer_default\n#SBATCH --requeue\n#SBATCH --signal=b:usr1@300 # 5 min before timeout\n\n# Log the sbatch script\ncat $0\n\nmodule unload mpi/openmpi/5.0\nmodule unload devel/cuda/12.4\nsource .venv/bin/activate\n\narray_records_dir_train=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_coinrun/coinrun_episodes_10m_gt_actions_split/train\narray_records_dir_val=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/data_coinrun/coinrun_episodes_10m_gt_actions_split/val\n\n\njob_name=$SLURM_JOB_NAME\nslurm_job_id=$SLURM_JOB_ID\n\nCHECKPOINT_DIR=/hkfs/work/workspace/scratch/tum_ind3695-jafa_ws_shared/checkpoints/big-runs/tokenizer/$job_name/$slurm_job_id\nmkdir -p $CHECKPOINT_DIR\n\nenv | grep SLURM\n\nsrun python jasmine/train_tokenizer.py \\n    --save_ckpt \\n    --image_height=64 \\n    --image_width=64 \\n    --ckpt_dir $CHECKPOINT_DIR \\n    --batch_size=120 \\n    --init_lr=0 \\n    --patch_size=16 \\n    --log \\n    --name=coinrun-tokenizer-default-$slurm_job_id \\n    --tags tokenizer coinrun default \\n    --entity instant-uv \\n    --project jafar \\n    --data_dir $array_records_dir_train \\n    --val_data_dir $array_records_dir_val \\n",shellscript,tab
+133,198803,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh",1326,0,"",shellscript,selection_mouse
+134,201451,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh",1310,22,"",shellscript,content
+135,201493,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh",1314,0,"",shellscript,selection_command
+136,205494,"TERMINAL",0,0,"sync-runner",,terminal_command
+137,205568,"TERMINAL",0,0,"]633;Csending incremental file list\r\n",,terminal_output
+138,207803,"TERMINAL",0,0,"slurm/jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh\r\n",,terminal_output
+139,207875,"TERMINAL",0,0,"\r\nsent 38,779 bytes  received 260 bytes  15,615.60 bytes/sec\r\ntotal size is 28,210,903  speedup is 722.63\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine_jobs",,terminal_output
+140,209763,"TERMINAL",0,0,"runner",,terminal_command
+141,213567,"TERMINAL",0,0,"sbatch slurm/jobs/mihir/horeka/coinrun/default_runs/train_tokenizer_default.sh",,terminal_command
+142,213576,"TERMINAL",0,0,"]633;CSubmitted batch job 3529698\r\n]0;tum_cte0515@hkn1993:~/Projects/jasmine_jobs",,terminal_output
+143,224381,"TERMINAL",0,0,"queue",,terminal_command
+144,224458,"TERMINAL",0,0,"]633;C[?1049h(B[?7hEvery 1.0s: squeue --mehkn1993.localdomain: Mon Sep 29 20:34:28 2025JOBID PARTITION     NAME     USER ST\tTIME  NODES NODELIST(REASON)3528968 accelerat train_la tum_cte0 PD\t0:00\t  8 (Priority)3528969 accelerat train_to tum_cte0 PD\t0:00\t  8 (Priority)3529698 accelerat train_to tum_cte0 PD\t0:00\t  1 (Priority)3529697 accelerat train_la tum_cte0  R\t0:36\t  1 hkn04293528955 accelerat train_to tum_cte0  R    7:29:49\t  1 hkn0709",,terminal_output
+145,225587,"TERMINAL",0,0,"9750",,terminal_output
+146,226585,"TERMINAL",0,0,"3081",,terminal_output
+147,227582,"TERMINAL",0,0,"192",,terminal_output
+148,228632,"TERMINAL",0,0,"2403",,terminal_output
+149,229760,"TERMINAL",0,0,"314",,terminal_output
+150,230887,"TERMINAL",0,0,"425",,terminal_output
+151,231807,"TERMINAL",0,0,"536",,terminal_output
+152,232814,"TERMINAL",0,0,"647",,terminal_output
+153,233981,"TERMINAL",0,0,"758",,terminal_output
+154,234830,"TERMINAL",0,0,"[?1049l\r[?1l>]0;tum_cte0515@hkn1993:~/Projects/jasmine_jobs",,terminal_output