Add files using upload-large-folder tool

healed/keep50_warmup_fixed_s1224/step0050/chat_template.jinja

@@ -0,0 +1,9 @@
+{{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'user' %}{{ '<|user|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
+'  + message['content'] + eos_token + '
+' }}{% else %}{{ '<|assistant|>
+'  + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
+' }}{% endif %}{% endfor %}

healed/keep50_warmup_fixed_s1224/step0050/config.json

@@ -0,0 +1,887 @@
+{
+  "architectures": [
+    "PrunedOlmoeForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
+    "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
+  },
+  "clip_qkv": null,
+  "dtype": "bfloat16",
+  "eos_token_id": 50279,
+  "expert_widths": [
+    [
+      1024,
+      384,
+      256,
+      256,
+      768,
+      1024,
+      128,
+      768,
+      384,
+      128,
+      1024,
+      384,
+      384,
+      128,
+      640,
+      896,
+      896,
+      256,
+      768,
+      128,
+      768,
+      768,
+      768,
+      768,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      256,
+      768,
+      512,
+      640,
+      768,
+      896,
+      1024,
+      384,
+      384,
+      512,
+      384,
+      256,
+      768,
+      640,
+      896,
+      896,
+      512,
+      768,
+      512,
+      384,
+      1024,
+      896,
+      896,
+      768,
+      128
+    ],
+    [
+      768,
+      896,
+      896,
+      256,
+      896,
+      384,
+      768,
+      768,
+      1024,
+      384,
+      512,
+      1024,
+      768,
+      768,
+      1024,
+      640,
+      128,
+      768,
+      384,
+      768,
+      128,
+      256,
+      512,
+      384,
+      896,
+      256,
+      1024,
+      1024,
+      896,
+      1024,
+      896,
+      768,
+      256,
+      768,
+      128,
+      1024,
+      768,
+      768,
+      640,
+      384,
+      1024,
+      384,
+      640,
+      896,
+      1024,
+      512,
+      512,
+      512,
+      640
+    ],
+    [
+      768,
+      1024,
+      512,
+      384,
+      384,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      640,
+      256,
+      1024,
+      1024,
+      1024,
+      640,
+      1024,
+      896,
+      512,
+      768,
+      128,
+      768,
+      1024,
+      768,
+      256,
+      1024,
+      896,
+      1024,
+      896,
+      512,
+      640,
+      384,
+      896,
+      384,
+      512,
+      640,
+      256,
+      768,
+      256,
+      768,
+      768,
+      640,
+      1024,
+      640,
+      896,
+      768,
+      768,
+      256
+    ],
+    [
+      640,
+      640,
+      896,
+      512,
+      640,
+      896,
+      896,
+      384,
+      896,
+      896,
+      384,
+      896,
+      640,
+      256,
+      640,
+      1024,
+      1024,
+      768,
+      1024,
+      896,
+      1024,
+      768,
+      896,
+      256,
+      512,
+      768,
+      1024,
+      256,
+      768,
+      512,
+      256,
+      640,
+      1024,
+      1024,
+      512,
+      1024,
+      768,
+      256,
+      1024,
+      384,
+      896,
+      1024,
+      896,
+      1024,
+      384
+    ],
+    [
+      256,
+      640,
+      640,
+      896,
+      768,
+      896,
+      768,
+      768,
+      896,
+      896,
+      1024,
+      256,
+      512,
+      1024,
+      640,
+      896,
+      512,
+      512,
+      384,
+      384,
+      256,
+      384,
+      384,
+      896,
+      896,
+      768,
+      640,
+      896,
+      768,
+      1024,
+      512,
+      640,
+      512,
+      640,
+      896,
+      512,
+      512,
+      384,
+      640,
+      896,
+      896,
+      896,
+      1024,
+      640,
+      1024,
+      640,
+      1024
+    ],
+    [
+      1024,
+      512,
+      1024,
+      1024,
+      640,
+      896,
+      640,
+      1024,
+      896,
+      384,
+      1024,
+      128,
+      896,
+      768,
+      1024,
+      768,
+      640,
+      896,
+      768,
+      640,
+      512,
+      896,
+      512,
+      640,
+      256,
+      768,
+      640,
+      768,
+      384,
+      896,
+      512,
+      512,
+      256,
+      512,
+      896,
+      256,
+      384,
+      640,
+      512,
+      640,
+      896,
+      512,
+      1024,
+      256,
+      768,
+      1024,
+      768,
+      256,
+      256
+    ],
+    [
+      640,
+      896,
+      1024,
+      896,
+      1024,
+      1024,
+      1024,
+      512,
+      256,
+      256,
+      1024,
+      768,
+      512,
+      768,
+      1024,
+      1024,
+      1024,
+      384,
+      512,
+      1024,
+      512,
+      1024,
+      128,
+      640,
+      640,
+      896,
+      768,
+      128,
+      256,
+      256,
+      256,
+      256,
+      896,
+      1024,
+      1024,
+      384,
+      896,
+      256,
+      896,
+      640,
+      1024,
+      384,
+      640,
+      256,
+      1024,
+      1024,
+      1024
+    ],
+    [
+      1024,
+      384,
+      1024,
+      1024,
+      256,
+      128,
+      256,
+      384,
+      256,
+      384,
+      896,
+      768,
+      896,
+      896,
+      512,
+      896,
+      640,
+      384,
+      384,
+      896,
+      768,
+      384,
+      896,
+      768,
+      768,
+      512,
+      896,
+      768,
+      768,
+      896,
+      128,
+      896,
+      512,
+      256,
+      768,
+      128,
+      384,
+      256,
+      896,
+      896,
+      384,
+      768,
+      512,
+      640,
+      256,
+      768,
+      640,
+      896,
+      384,
+      512,
+      1024,
+      768,
+      384
+    ],
+    [
+      512,
+      768,
+      512,
+      256,
+      128,
+      640,
+      384,
+      640,
+      768,
+      896,
+      640,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      256,
+      896,
+      512,
+      256,
+      896,
+      768,
+      256,
+      896,
+      896,
+      384,
+      896,
+      640,
+      768,
+      512,
+      768,
+      768,
+      1024,
+      768,
+      640,
+      768,
+      384,
+      256,
+      512,
+      896,
+      128,
+      384,
+      256,
+      768,
+      384,
+      256,
+      1024,
+      1024,
+      896,
+      256,
+      1024,
+      256,
+      128,
+      896
+    ],
+    [
+      640,
+      640,
+      896,
+      256,
+      1024,
+      512,
+      1024,
+      768,
+      384,
+      512,
+      256,
+      768,
+      896,
+      768,
+      512,
+      768,
+      768,
+      640,
+      384,
+      768,
+      512,
+      768,
+      768,
+      512,
+      768,
+      128,
+      896,
+      512,
+      768,
+      1024,
+      128,
+      384,
+      768,
+      768,
+      768,
+      384,
+      512,
+      640,
+      768,
+      512,
+      768,
+      1024,
+      640,
+      896,
+      256,
+      1024,
+      384,
+      768,
+      768,
+      768
+    ],
+    [
+      896,
+      512,
+      896,
+      768,
+      384,
+      384,
+      768,
+      512,
+      768,
+      512,
+      1024,
+      640,
+      896,
+      896,
+      256,
+      640,
+      1024,
+      256,
+      896,
+      128,
+      128,
+      128,
+      768,
+      896,
+      384,
+      896,
+      512,
+      896,
+      384,
+      256,
+      640,
+      640,
+      896,
+      768,
+      640,
+      256,
+      896,
+      896,
+      512,
+      128,
+      896,
+      256,
+      256,
+      640,
+      896,
+      896,
+      128,
+      1024,
+      256,
+      384,
+      1024,
+      640,
+      896
+    ],
+    [
+      768,
+      384,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      128,
+      128,
+      896,
+      256,
+      384,
+      896,
+      512,
+      768,
+      768,
+      512,
+      512,
+      768,
+      896,
+      640,
+      768,
+      896,
+      896,
+      896,
+      640,
+      896,
+      640,
+      512,
+      896,
+      256,
+      512,
+      128,
+      512,
+      384,
+      768,
+      768,
+      1024,
+      256,
+      768,
+      256,
+      768,
+      512,
+      640,
+      1024,
+      128,
+      896,
+      896,
+      896,
+      896,
+      1024
+    ],
+    [
+      512,
+      384,
+      768,
+      640,
+      640,
+      768,
+      1024,
+      896,
+      512,
+      256,
+      640,
+      768,
+      640,
+      896,
+      128,
+      256,
+      896,
+      1024,
+      256,
+      640,
+      512,
+      256,
+      128,
+      512,
+      256,
+      640,
+      768,
+      768,
+      128,
+      128,
+      768,
+      640,
+      1024,
+      1024,
+      768,
+      512,
+      896,
+      768,
+      896,
+      1024,
+      896,
+      896,
+      512,
+      640,
+      1024,
+      512,
+      1024,
+      512,
+      512,
+      512,
+      768
+    ],
+    [
+      896,
+      896,
+      1024,
+      1024,
+      896,
+      128,
+      768,
+      256,
+      1024,
+      256,
+      1024,
+      640,
+      384,
+      256,
+      256,
+      512,
+      768,
+      896,
+      512,
+      768,
+      384,
+      1024,
+      896,
+      896,
+      1024,
+      896,
+      768,
+      896,
+      640,
+      1024,
+      512,
+      896,
+      512,
+      1024,
+      512,
+      512,
+      256,
+      256,
+      256,
+      512,
+      768,
+      128,
+      384,
+      512,
+      896,
+      896,
+      1024,
+      256,
+      384,
+      640
+    ],
+    [
+      896,
+      640,
+      384,
+      512,
+      256,
+      640,
+      1024,
+      384,
+      1024,
+      1024,
+      768,
+      256,
+      1024,
+      768,
+      512,
+      896,
+      256,
+      1024,
+      768,
+      768,
+      768,
+      384,
+      384,
+      256,
+      1024,
+      384,
+      384,
+      384,
+      896,
+      768,
+      640,
+      768,
+      512,
+      896,
+      896,
+      896,
+      896,
+      256,
+      384,
+      128,
+      1024,
+      896,
+      256,
+      256,
+      768,
+      640,
+      896,
+      384,
+      768,
+      512,
+      640
+    ],
+    [
+      896,
+      1024,
+      768,
+      1024,
+      896,
+      256,
+      768,
+      128,
+      128,
+      768,
+      512,
+      896,
+      384,
+      768,
+      1024,
+      256,
+      768,
+      768,
+      256,
+      512,
+      512,
+      640,
+      512,
+      256,
+      768,
+      896,
+      384,
+      1024,
+      640,
+      1024,
+      512,
+      512,
+      384,
+      512,
+      512,
+      1024,
+      384,
+      896,
+      768,
+      384,
+      384,
+      128,
+      384,
+      1024,
+      896,
+      640,
+      768,
+      768,
+      256,
+      640,
+      512,
+      640,
+      384
+    ]
+  ],
+  "glean_metadata": {
+    "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
+    "block_size": 128,
+    "criterion": "reap",
+    "dead_experts": 217,
+    "keep_fraction": 0.5,
+    "min_width": 128,
+    "params": 3697491968,
+    "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
+  },
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 1024,
+  "max_position_embeddings": 4096,
+  "model_type": "pruned_olmoe",
+  "norm_topk_prob": false,
+  "num_attention_heads": 16,
+  "num_experts": 64,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "output_router_logits": false,
+  "pad_token_id": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "router_aux_loss_coef": 0.01,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6",
+  "use_cache": false,
+  "vocab_size": 50304
+}

healed/keep50_warmup_fixed_s1224/step0050/configuration_pruned_olmoe.py

@@ -0,0 +1,27 @@
+"""Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
+"""
+
+from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
+
+
+class PrunedOlmoeConfig(OlmoeConfig):
+    """OlmoeConfig plus a per-(layer, expert) width table.
+
+    ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
+    expert in decoder layer ``l``, in expert order. Lists are ragged: layers
+    may keep different numbers of experts (deleted experts simply don't
+    appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
+    and each width may differ (multiples of the GEMM block size, 128, for
+    variable-MegaBlocks execution). ``None`` means an unpruned model
+    (uniform ``num_experts`` × ``intermediate_size``).
+
+    The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
+    (pre-pruning) values for provenance; the width table is authoritative for
+    the built architecture.
+    """
+
+    model_type = "pruned_olmoe"
+
+    def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
+        super().__init__(**kwargs)
+        self.expert_widths = expert_widths

healed/keep50_warmup_fixed_s1224/step0050/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 50279,
+  "pad_token_id": 1,
+  "transformers_version": "4.57.6"
+}

healed/keep50_warmup_fixed_s1224/step0050/modeling_pruned_olmoe.py

@@ -0,0 +1,66 @@
+"""GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
+
+Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
+(docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
+per-(layer, expert) width table: ``super().__init__`` builds the uniform
+architecture from the config, then every MoE block is rebuilt to its pruned
+shape — surviving experts only, each at its own width, router sliced to
+match — so the state dict aligns exactly with what
+``glean.prune.prune_channels_global`` leaves behind.
+
+Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
+uniform ``config.num_experts`` and is unsupported on ragged models.
+"""
+
+import torch.nn as nn
+from transformers.activations import ACT2FN
+from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
+
+from .configuration_pruned_olmoe import PrunedOlmoeConfig
+
+
+class RaggedOlmoeMLP(nn.Module):
+    """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
+
+    def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
+    """OLMoE with per-layer surviving-expert lists at per-expert widths."""
+
+    config_class = PrunedOlmoeConfig
+
+    def __init__(self, config: PrunedOlmoeConfig):
+        super().__init__(config)
+        widths_table = getattr(config, "expert_widths", None)
+        if widths_table is None:
+            return  # unpruned: plain OLMoE
+        if len(widths_table) != len(self.model.layers):
+            raise ValueError(
+                f"expert_widths has {len(widths_table)} rows but the model has "
+                f"{len(self.model.layers)} decoder layers"
+            )
+        for layer, widths in zip(self.model.layers, widths_table):
+            if any(w <= 0 for w in widths):
+                raise ValueError("expert_widths must list surviving experts only (>0)")
+            block = layer.mlp
+            if len(widths) < block.top_k:
+                raise ValueError(
+                    f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
+                )
+            block.num_experts = len(widths)
+            block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
+            block.experts = nn.ModuleList(
+                RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
+                for w in widths
+            )

healed/keep50_warmup_fixed_s1224/step0050/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "|||IP_ADDRESS|||",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "|||IP_ADDRESS|||",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

healed/keep50_warmup_fixed_s1224/step0050/tokenizer_config.json

@@ -0,0 +1,247 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50254": {
+      "content": "                        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50255": {
+      "content": "                       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50256": {
+      "content": "                      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50257": {
+      "content": "                     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50258": {
+      "content": "                    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50259": {
+      "content": "                   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50260": {
+      "content": "                  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50261": {
+      "content": "                 ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50262": {
+      "content": "                ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50263": {
+      "content": "               ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50264": {
+      "content": "              ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50265": {
+      "content": "             ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50266": {
+      "content": "            ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50267": {
+      "content": "           ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50268": {
+      "content": "          ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50269": {
+      "content": "         ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50270": {
+      "content": "        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50271": {
+      "content": "       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50272": {
+      "content": "      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50273": {
+      "content": "     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50274": {
+      "content": "    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50275": {
+      "content": "   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50276": {
+      "content": "  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50277": {
+      "content": "|||EMAIL_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50278": {
+      "content": "|||PHONE_NUMBER|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50279": {
+      "content": "|||IP_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50280": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "|||IP_ADDRESS|||",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "|||IP_ADDRESS|||",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": null
+}

healed/keep50_warmup_fixed_s1224/step0100/chat_template.jinja

@@ -0,0 +1,9 @@
+{{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'user' %}{{ '<|user|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
+'  + message['content'] + eos_token + '
+' }}{% else %}{{ '<|assistant|>
+'  + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
+' }}{% endif %}{% endfor %}

healed/keep50_warmup_fixed_s1224/step0100/config.json

@@ -0,0 +1,887 @@
+{
+  "architectures": [
+    "PrunedOlmoeForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
+    "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
+  },
+  "clip_qkv": null,
+  "dtype": "bfloat16",
+  "eos_token_id": 50279,
+  "expert_widths": [
+    [
+      1024,
+      384,
+      256,
+      256,
+      768,
+      1024,
+      128,
+      768,
+      384,
+      128,
+      1024,
+      384,
+      384,
+      128,
+      640,
+      896,
+      896,
+      256,
+      768,
+      128,
+      768,
+      768,
+      768,
+      768,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      256,
+      768,
+      512,
+      640,
+      768,
+      896,
+      1024,
+      384,
+      384,
+      512,
+      384,
+      256,
+      768,
+      640,
+      896,
+      896,
+      512,
+      768,
+      512,
+      384,
+      1024,
+      896,
+      896,
+      768,
+      128
+    ],
+    [
+      768,
+      896,
+      896,
+      256,
+      896,
+      384,
+      768,
+      768,
+      1024,
+      384,
+      512,
+      1024,
+      768,
+      768,
+      1024,
+      640,
+      128,
+      768,
+      384,
+      768,
+      128,
+      256,
+      512,
+      384,
+      896,
+      256,
+      1024,
+      1024,
+      896,
+      1024,
+      896,
+      768,
+      256,
+      768,
+      128,
+      1024,
+      768,
+      768,
+      640,
+      384,
+      1024,
+      384,
+      640,
+      896,
+      1024,
+      512,
+      512,
+      512,
+      640
+    ],
+    [
+      768,
+      1024,
+      512,
+      384,
+      384,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      640,
+      256,
+      1024,
+      1024,
+      1024,
+      640,
+      1024,
+      896,
+      512,
+      768,
+      128,
+      768,
+      1024,
+      768,
+      256,
+      1024,
+      896,
+      1024,
+      896,
+      512,
+      640,
+      384,
+      896,
+      384,
+      512,
+      640,
+      256,
+      768,
+      256,
+      768,
+      768,
+      640,
+      1024,
+      640,
+      896,
+      768,
+      768,
+      256
+    ],
+    [
+      640,
+      640,
+      896,
+      512,
+      640,
+      896,
+      896,
+      384,
+      896,
+      896,
+      384,
+      896,
+      640,
+      256,
+      640,
+      1024,
+      1024,
+      768,
+      1024,
+      896,
+      1024,
+      768,
+      896,
+      256,
+      512,
+      768,
+      1024,
+      256,
+      768,
+      512,
+      256,
+      640,
+      1024,
+      1024,
+      512,
+      1024,
+      768,
+      256,
+      1024,
+      384,
+      896,
+      1024,
+      896,
+      1024,
+      384
+    ],
+    [
+      256,
+      640,
+      640,
+      896,
+      768,
+      896,
+      768,
+      768,
+      896,
+      896,
+      1024,
+      256,
+      512,
+      1024,
+      640,
+      896,
+      512,
+      512,
+      384,
+      384,
+      256,
+      384,
+      384,
+      896,
+      896,
+      768,
+      640,
+      896,
+      768,
+      1024,
+      512,
+      640,
+      512,
+      640,
+      896,
+      512,
+      512,
+      384,
+      640,
+      896,
+      896,
+      896,
+      1024,
+      640,
+      1024,
+      640,
+      1024
+    ],
+    [
+      1024,
+      512,
+      1024,
+      1024,
+      640,
+      896,
+      640,
+      1024,
+      896,
+      384,
+      1024,
+      128,
+      896,
+      768,
+      1024,
+      768,
+      640,
+      896,
+      768,
+      640,
+      512,
+      896,
+      512,
+      640,
+      256,
+      768,
+      640,
+      768,
+      384,
+      896,
+      512,
+      512,
+      256,
+      512,
+      896,
+      256,
+      384,
+      640,
+      512,
+      640,
+      896,
+      512,
+      1024,
+      256,
+      768,
+      1024,
+      768,
+      256,
+      256
+    ],
+    [
+      640,
+      896,
+      1024,
+      896,
+      1024,
+      1024,
+      1024,
+      512,
+      256,
+      256,
+      1024,
+      768,
+      512,
+      768,
+      1024,
+      1024,
+      1024,
+      384,
+      512,
+      1024,
+      512,
+      1024,
+      128,
+      640,
+      640,
+      896,
+      768,
+      128,
+      256,
+      256,
+      256,
+      256,
+      896,
+      1024,
+      1024,
+      384,
+      896,
+      256,
+      896,
+      640,
+      1024,
+      384,
+      640,
+      256,
+      1024,
+      1024,
+      1024
+    ],
+    [
+      1024,
+      384,
+      1024,
+      1024,
+      256,
+      128,
+      256,
+      384,
+      256,
+      384,
+      896,
+      768,
+      896,
+      896,
+      512,
+      896,
+      640,
+      384,
+      384,
+      896,
+      768,
+      384,
+      896,
+      768,
+      768,
+      512,
+      896,
+      768,
+      768,
+      896,
+      128,
+      896,
+      512,
+      256,
+      768,
+      128,
+      384,
+      256,
+      896,
+      896,
+      384,
+      768,
+      512,
+      640,
+      256,
+      768,
+      640,
+      896,
+      384,
+      512,
+      1024,
+      768,
+      384
+    ],
+    [
+      512,
+      768,
+      512,
+      256,
+      128,
+      640,
+      384,
+      640,
+      768,
+      896,
+      640,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      256,
+      896,
+      512,
+      256,
+      896,
+      768,
+      256,
+      896,
+      896,
+      384,
+      896,
+      640,
+      768,
+      512,
+      768,
+      768,
+      1024,
+      768,
+      640,
+      768,
+      384,
+      256,
+      512,
+      896,
+      128,
+      384,
+      256,
+      768,
+      384,
+      256,
+      1024,
+      1024,
+      896,
+      256,
+      1024,
+      256,
+      128,
+      896
+    ],
+    [
+      640,
+      640,
+      896,
+      256,
+      1024,
+      512,
+      1024,
+      768,
+      384,
+      512,
+      256,
+      768,
+      896,
+      768,
+      512,
+      768,
+      768,
+      640,
+      384,
+      768,
+      512,
+      768,
+      768,
+      512,
+      768,
+      128,
+      896,
+      512,
+      768,
+      1024,
+      128,
+      384,
+      768,
+      768,
+      768,
+      384,
+      512,
+      640,
+      768,
+      512,
+      768,
+      1024,
+      640,
+      896,
+      256,
+      1024,
+      384,
+      768,
+      768,
+      768
+    ],
+    [
+      896,
+      512,
+      896,
+      768,
+      384,
+      384,
+      768,
+      512,
+      768,
+      512,
+      1024,
+      640,
+      896,
+      896,
+      256,
+      640,
+      1024,
+      256,
+      896,
+      128,
+      128,
+      128,
+      768,
+      896,
+      384,
+      896,
+      512,
+      896,
+      384,
+      256,
+      640,
+      640,
+      896,
+      768,
+      640,
+      256,
+      896,
+      896,
+      512,
+      128,
+      896,
+      256,
+      256,
+      640,
+      896,
+      896,
+      128,
+      1024,
+      256,
+      384,
+      1024,
+      640,
+      896
+    ],
+    [
+      768,
+      384,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      128,
+      128,
+      896,
+      256,
+      384,
+      896,
+      512,
+      768,
+      768,
+      512,
+      512,
+      768,
+      896,
+      640,
+      768,
+      896,
+      896,
+      896,
+      640,
+      896,
+      640,
+      512,
+      896,
+      256,
+      512,
+      128,
+      512,
+      384,
+      768,
+      768,
+      1024,
+      256,
+      768,
+      256,
+      768,
+      512,
+      640,
+      1024,
+      128,
+      896,
+      896,
+      896,
+      896,
+      1024
+    ],
+    [
+      512,
+      384,
+      768,
+      640,
+      640,
+      768,
+      1024,
+      896,
+      512,
+      256,
+      640,
+      768,
+      640,
+      896,
+      128,
+      256,
+      896,
+      1024,
+      256,
+      640,
+      512,
+      256,
+      128,
+      512,
+      256,
+      640,
+      768,
+      768,
+      128,
+      128,
+      768,
+      640,
+      1024,
+      1024,
+      768,
+      512,
+      896,
+      768,
+      896,
+      1024,
+      896,
+      896,
+      512,
+      640,
+      1024,
+      512,
+      1024,
+      512,
+      512,
+      512,
+      768
+    ],
+    [
+      896,
+      896,
+      1024,
+      1024,
+      896,
+      128,
+      768,
+      256,
+      1024,
+      256,
+      1024,
+      640,
+      384,
+      256,
+      256,
+      512,
+      768,
+      896,
+      512,
+      768,
+      384,
+      1024,
+      896,
+      896,
+      1024,
+      896,
+      768,
+      896,
+      640,
+      1024,
+      512,
+      896,
+      512,
+      1024,
+      512,
+      512,
+      256,
+      256,
+      256,
+      512,
+      768,
+      128,
+      384,
+      512,
+      896,
+      896,
+      1024,
+      256,
+      384,
+      640
+    ],
+    [
+      896,
+      640,
+      384,
+      512,
+      256,
+      640,
+      1024,
+      384,
+      1024,
+      1024,
+      768,
+      256,
+      1024,
+      768,
+      512,
+      896,
+      256,
+      1024,
+      768,
+      768,
+      768,
+      384,
+      384,
+      256,
+      1024,
+      384,
+      384,
+      384,
+      896,
+      768,
+      640,
+      768,
+      512,
+      896,
+      896,
+      896,
+      896,
+      256,
+      384,
+      128,
+      1024,
+      896,
+      256,
+      256,
+      768,
+      640,
+      896,
+      384,
+      768,
+      512,
+      640
+    ],
+    [
+      896,
+      1024,
+      768,
+      1024,
+      896,
+      256,
+      768,
+      128,
+      128,
+      768,
+      512,
+      896,
+      384,
+      768,
+      1024,
+      256,
+      768,
+      768,
+      256,
+      512,
+      512,
+      640,
+      512,
+      256,
+      768,
+      896,
+      384,
+      1024,
+      640,
+      1024,
+      512,
+      512,
+      384,
+      512,
+      512,
+      1024,
+      384,
+      896,
+      768,
+      384,
+      384,
+      128,
+      384,
+      1024,
+      896,
+      640,
+      768,
+      768,
+      256,
+      640,
+      512,
+      640,
+      384
+    ]
+  ],
+  "glean_metadata": {
+    "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
+    "block_size": 128,
+    "criterion": "reap",
+    "dead_experts": 217,
+    "keep_fraction": 0.5,
+    "min_width": 128,
+    "params": 3697491968,
+    "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
+  },
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 1024,
+  "max_position_embeddings": 4096,
+  "model_type": "pruned_olmoe",
+  "norm_topk_prob": false,
+  "num_attention_heads": 16,
+  "num_experts": 64,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "output_router_logits": false,
+  "pad_token_id": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "router_aux_loss_coef": 0.01,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6",
+  "use_cache": false,
+  "vocab_size": 50304
+}

healed/keep50_warmup_fixed_s1224/step0100/configuration_pruned_olmoe.py

@@ -0,0 +1,27 @@
+"""Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
+"""
+
+from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
+
+
+class PrunedOlmoeConfig(OlmoeConfig):
+    """OlmoeConfig plus a per-(layer, expert) width table.
+
+    ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
+    expert in decoder layer ``l``, in expert order. Lists are ragged: layers
+    may keep different numbers of experts (deleted experts simply don't
+    appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
+    and each width may differ (multiples of the GEMM block size, 128, for
+    variable-MegaBlocks execution). ``None`` means an unpruned model
+    (uniform ``num_experts`` × ``intermediate_size``).
+
+    The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
+    (pre-pruning) values for provenance; the width table is authoritative for
+    the built architecture.
+    """
+
+    model_type = "pruned_olmoe"
+
+    def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
+        super().__init__(**kwargs)
+        self.expert_widths = expert_widths

healed/keep50_warmup_fixed_s1224/step0100/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 50279,
+  "pad_token_id": 1,
+  "transformers_version": "4.57.6"
+}

healed/keep50_warmup_fixed_s1224/step0100/modeling_pruned_olmoe.py

@@ -0,0 +1,66 @@
+"""GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
+
+Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
+(docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
+per-(layer, expert) width table: ``super().__init__`` builds the uniform
+architecture from the config, then every MoE block is rebuilt to its pruned
+shape — surviving experts only, each at its own width, router sliced to
+match — so the state dict aligns exactly with what
+``glean.prune.prune_channels_global`` leaves behind.
+
+Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
+uniform ``config.num_experts`` and is unsupported on ragged models.
+"""
+
+import torch.nn as nn
+from transformers.activations import ACT2FN
+from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
+
+from .configuration_pruned_olmoe import PrunedOlmoeConfig
+
+
+class RaggedOlmoeMLP(nn.Module):
+    """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
+
+    def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
+    """OLMoE with per-layer surviving-expert lists at per-expert widths."""
+
+    config_class = PrunedOlmoeConfig
+
+    def __init__(self, config: PrunedOlmoeConfig):
+        super().__init__(config)
+        widths_table = getattr(config, "expert_widths", None)
+        if widths_table is None:
+            return  # unpruned: plain OLMoE
+        if len(widths_table) != len(self.model.layers):
+            raise ValueError(
+                f"expert_widths has {len(widths_table)} rows but the model has "
+                f"{len(self.model.layers)} decoder layers"
+            )
+        for layer, widths in zip(self.model.layers, widths_table):
+            if any(w <= 0 for w in widths):
+                raise ValueError("expert_widths must list surviving experts only (>0)")
+            block = layer.mlp
+            if len(widths) < block.top_k:
+                raise ValueError(
+                    f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
+                )
+            block.num_experts = len(widths)
+            block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
+            block.experts = nn.ModuleList(
+                RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
+                for w in widths
+            )

healed/keep50_warmup_fixed_s1224/step0100/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "|||IP_ADDRESS|||",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "|||IP_ADDRESS|||",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

healed/keep50_warmup_fixed_s1224/step0100/tokenizer_config.json

@@ -0,0 +1,247 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50254": {
+      "content": "                        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50255": {
+      "content": "                       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50256": {
+      "content": "                      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50257": {
+      "content": "                     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50258": {
+      "content": "                    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50259": {
+      "content": "                   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50260": {
+      "content": "                  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50261": {
+      "content": "                 ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50262": {
+      "content": "                ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50263": {
+      "content": "               ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50264": {
+      "content": "              ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50265": {
+      "content": "             ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50266": {
+      "content": "            ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50267": {
+      "content": "           ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50268": {
+      "content": "          ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50269": {
+      "content": "         ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50270": {
+      "content": "        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50271": {
+      "content": "       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50272": {
+      "content": "      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50273": {
+      "content": "     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50274": {
+      "content": "    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50275": {
+      "content": "   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50276": {
+      "content": "  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50277": {
+      "content": "|||EMAIL_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50278": {
+      "content": "|||PHONE_NUMBER|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50279": {
+      "content": "|||IP_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50280": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "|||IP_ADDRESS|||",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "|||IP_ADDRESS|||",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": null
+}

healed/mixceonly_keep50/vllm_live/chat_template.jinja

@@ -0,0 +1,9 @@
+{{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'user' %}{{ '<|user|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
+'  + message['content'] + eos_token + '
+' }}{% else %}{{ '<|assistant|>
+'  + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
+' }}{% endif %}{% endfor %}

healed/mixceonly_keep50/vllm_live/config.json

@@ -0,0 +1,887 @@
+{
+  "architectures": [
+    "PrunedOlmoeForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
+    "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
+  },
+  "clip_qkv": null,
+  "dtype": "bfloat16",
+  "eos_token_id": 50279,
+  "expert_widths": [
+    [
+      1024,
+      384,
+      256,
+      256,
+      768,
+      1024,
+      128,
+      768,
+      384,
+      128,
+      1024,
+      384,
+      384,
+      128,
+      640,
+      896,
+      896,
+      256,
+      768,
+      128,
+      768,
+      768,
+      768,
+      768,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      256,
+      768,
+      512,
+      640,
+      768,
+      896,
+      1024,
+      384,
+      384,
+      512,
+      384,
+      256,
+      768,
+      640,
+      896,
+      896,
+      512,
+      768,
+      512,
+      384,
+      1024,
+      896,
+      896,
+      768,
+      128
+    ],
+    [
+      768,
+      896,
+      896,
+      256,
+      896,
+      384,
+      768,
+      768,
+      1024,
+      384,
+      512,
+      1024,
+      768,
+      768,
+      1024,
+      640,
+      128,
+      768,
+      384,
+      768,
+      128,
+      256,
+      512,
+      384,
+      896,
+      256,
+      1024,
+      1024,
+      896,
+      1024,
+      896,
+      768,
+      256,
+      768,
+      128,
+      1024,
+      768,
+      768,
+      640,
+      384,
+      1024,
+      384,
+      640,
+      896,
+      1024,
+      512,
+      512,
+      512,
+      640
+    ],
+    [
+      768,
+      1024,
+      512,
+      384,
+      384,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      640,
+      256,
+      1024,
+      1024,
+      1024,
+      640,
+      1024,
+      896,
+      512,
+      768,
+      128,
+      768,
+      1024,
+      768,
+      256,
+      1024,
+      896,
+      1024,
+      896,
+      512,
+      640,
+      384,
+      896,
+      384,
+      512,
+      640,
+      256,
+      768,
+      256,
+      768,
+      768,
+      640,
+      1024,
+      640,
+      896,
+      768,
+      768,
+      256
+    ],
+    [
+      640,
+      640,
+      896,
+      512,
+      640,
+      896,
+      896,
+      384,
+      896,
+      896,
+      384,
+      896,
+      640,
+      256,
+      640,
+      1024,
+      1024,
+      768,
+      1024,
+      896,
+      1024,
+      768,
+      896,
+      256,
+      512,
+      768,
+      1024,
+      256,
+      768,
+      512,
+      256,
+      640,
+      1024,
+      1024,
+      512,
+      1024,
+      768,
+      256,
+      1024,
+      384,
+      896,
+      1024,
+      896,
+      1024,
+      384
+    ],
+    [
+      256,
+      640,
+      640,
+      896,
+      768,
+      896,
+      768,
+      768,
+      896,
+      896,
+      1024,
+      256,
+      512,
+      1024,
+      640,
+      896,
+      512,
+      512,
+      384,
+      384,
+      256,
+      384,
+      384,
+      896,
+      896,
+      768,
+      640,
+      896,
+      768,
+      1024,
+      512,
+      640,
+      512,
+      640,
+      896,
+      512,
+      512,
+      384,
+      640,
+      896,
+      896,
+      896,
+      1024,
+      640,
+      1024,
+      640,
+      1024
+    ],
+    [
+      1024,
+      512,
+      1024,
+      1024,
+      640,
+      896,
+      640,
+      1024,
+      896,
+      384,
+      1024,
+      128,
+      896,
+      768,
+      1024,
+      768,
+      640,
+      896,
+      768,
+      640,
+      512,
+      896,
+      512,
+      640,
+      256,
+      768,
+      640,
+      768,
+      384,
+      896,
+      512,
+      512,
+      256,
+      512,
+      896,
+      256,
+      384,
+      640,
+      512,
+      640,
+      896,
+      512,
+      1024,
+      256,
+      768,
+      1024,
+      768,
+      256,
+      256
+    ],
+    [
+      640,
+      896,
+      1024,
+      896,
+      1024,
+      1024,
+      1024,
+      512,
+      256,
+      256,
+      1024,
+      768,
+      512,
+      768,
+      1024,
+      1024,
+      1024,
+      384,
+      512,
+      1024,
+      512,
+      1024,
+      128,
+      640,
+      640,
+      896,
+      768,
+      128,
+      256,
+      256,
+      256,
+      256,
+      896,
+      1024,
+      1024,
+      384,
+      896,
+      256,
+      896,
+      640,
+      1024,
+      384,
+      640,
+      256,
+      1024,
+      1024,
+      1024
+    ],
+    [
+      1024,
+      384,
+      1024,
+      1024,
+      256,
+      128,
+      256,
+      384,
+      256,
+      384,
+      896,
+      768,
+      896,
+      896,
+      512,
+      896,
+      640,
+      384,
+      384,
+      896,
+      768,
+      384,
+      896,
+      768,
+      768,
+      512,
+      896,
+      768,
+      768,
+      896,
+      128,
+      896,
+      512,
+      256,
+      768,
+      128,
+      384,
+      256,
+      896,
+      896,
+      384,
+      768,
+      512,
+      640,
+      256,
+      768,
+      640,
+      896,
+      384,
+      512,
+      1024,
+      768,
+      384
+    ],
+    [
+      512,
+      768,
+      512,
+      256,
+      128,
+      640,
+      384,
+      640,
+      768,
+      896,
+      640,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      256,
+      896,
+      512,
+      256,
+      896,
+      768,
+      256,
+      896,
+      896,
+      384,
+      896,
+      640,
+      768,
+      512,
+      768,
+      768,
+      1024,
+      768,
+      640,
+      768,
+      384,
+      256,
+      512,
+      896,
+      128,
+      384,
+      256,
+      768,
+      384,
+      256,
+      1024,
+      1024,
+      896,
+      256,
+      1024,
+      256,
+      128,
+      896
+    ],
+    [
+      640,
+      640,
+      896,
+      256,
+      1024,
+      512,
+      1024,
+      768,
+      384,
+      512,
+      256,
+      768,
+      896,
+      768,
+      512,
+      768,
+      768,
+      640,
+      384,
+      768,
+      512,
+      768,
+      768,
+      512,
+      768,
+      128,
+      896,
+      512,
+      768,
+      1024,
+      128,
+      384,
+      768,
+      768,
+      768,
+      384,
+      512,
+      640,
+      768,
+      512,
+      768,
+      1024,
+      640,
+      896,
+      256,
+      1024,
+      384,
+      768,
+      768,
+      768
+    ],
+    [
+      896,
+      512,
+      896,
+      768,
+      384,
+      384,
+      768,
+      512,
+      768,
+      512,
+      1024,
+      640,
+      896,
+      896,
+      256,
+      640,
+      1024,
+      256,
+      896,
+      128,
+      128,
+      128,
+      768,
+      896,
+      384,
+      896,
+      512,
+      896,
+      384,
+      256,
+      640,
+      640,
+      896,
+      768,
+      640,
+      256,
+      896,
+      896,
+      512,
+      128,
+      896,
+      256,
+      256,
+      640,
+      896,
+      896,
+      128,
+      1024,
+      256,
+      384,
+      1024,
+      640,
+      896
+    ],
+    [
+      768,
+      384,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      128,
+      128,
+      896,
+      256,
+      384,
+      896,
+      512,
+      768,
+      768,
+      512,
+      512,
+      768,
+      896,
+      640,
+      768,
+      896,
+      896,
+      896,
+      640,
+      896,
+      640,
+      512,
+      896,
+      256,
+      512,
+      128,
+      512,
+      384,
+      768,
+      768,
+      1024,
+      256,
+      768,
+      256,
+      768,
+      512,
+      640,
+      1024,
+      128,
+      896,
+      896,
+      896,
+      896,
+      1024
+    ],
+    [
+      512,
+      384,
+      768,
+      640,
+      640,
+      768,
+      1024,
+      896,
+      512,
+      256,
+      640,
+      768,
+      640,
+      896,
+      128,
+      256,
+      896,
+      1024,
+      256,
+      640,
+      512,
+      256,
+      128,
+      512,
+      256,
+      640,
+      768,
+      768,
+      128,
+      128,
+      768,
+      640,
+      1024,
+      1024,
+      768,
+      512,
+      896,
+      768,
+      896,
+      1024,
+      896,
+      896,
+      512,
+      640,
+      1024,
+      512,
+      1024,
+      512,
+      512,
+      512,
+      768
+    ],
+    [
+      896,
+      896,
+      1024,
+      1024,
+      896,
+      128,
+      768,
+      256,
+      1024,
+      256,
+      1024,
+      640,
+      384,
+      256,
+      256,
+      512,
+      768,
+      896,
+      512,
+      768,
+      384,
+      1024,
+      896,
+      896,
+      1024,
+      896,
+      768,
+      896,
+      640,
+      1024,
+      512,
+      896,
+      512,
+      1024,
+      512,
+      512,
+      256,
+      256,
+      256,
+      512,
+      768,
+      128,
+      384,
+      512,
+      896,
+      896,
+      1024,
+      256,
+      384,
+      640
+    ],
+    [
+      896,
+      640,
+      384,
+      512,
+      256,
+      640,
+      1024,
+      384,
+      1024,
+      1024,
+      768,
+      256,
+      1024,
+      768,
+      512,
+      896,
+      256,
+      1024,
+      768,
+      768,
+      768,
+      384,
+      384,
+      256,
+      1024,
+      384,
+      384,
+      384,
+      896,
+      768,
+      640,
+      768,
+      512,
+      896,
+      896,
+      896,
+      896,
+      256,
+      384,
+      128,
+      1024,
+      896,
+      256,
+      256,
+      768,
+      640,
+      896,
+      384,
+      768,
+      512,
+      640
+    ],
+    [
+      896,
+      1024,
+      768,
+      1024,
+      896,
+      256,
+      768,
+      128,
+      128,
+      768,
+      512,
+      896,
+      384,
+      768,
+      1024,
+      256,
+      768,
+      768,
+      256,
+      512,
+      512,
+      640,
+      512,
+      256,
+      768,
+      896,
+      384,
+      1024,
+      640,
+      1024,
+      512,
+      512,
+      384,
+      512,
+      512,
+      1024,
+      384,
+      896,
+      768,
+      384,
+      384,
+      128,
+      384,
+      1024,
+      896,
+      640,
+      768,
+      768,
+      256,
+      640,
+      512,
+      640,
+      384
+    ]
+  ],
+  "glean_metadata": {
+    "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
+    "block_size": 128,
+    "criterion": "reap",
+    "dead_experts": 217,
+    "keep_fraction": 0.5,
+    "min_width": 128,
+    "params": 3697491968,
+    "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
+  },
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 1024,
+  "max_position_embeddings": 4096,
+  "model_type": "pruned_olmoe",
+  "norm_topk_prob": false,
+  "num_attention_heads": 16,
+  "num_experts": 64,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "output_router_logits": false,
+  "pad_token_id": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "router_aux_loss_coef": 0.01,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6",
+  "use_cache": false,
+  "vocab_size": 50304
+}

healed/mixceonly_keep50/vllm_live/configuration_pruned_olmoe.py

@@ -0,0 +1,27 @@
+"""Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
+"""
+
+from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
+
+
+class PrunedOlmoeConfig(OlmoeConfig):
+    """OlmoeConfig plus a per-(layer, expert) width table.
+
+    ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
+    expert in decoder layer ``l``, in expert order. Lists are ragged: layers
+    may keep different numbers of experts (deleted experts simply don't
+    appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
+    and each width may differ (multiples of the GEMM block size, 128, for
+    variable-MegaBlocks execution). ``None`` means an unpruned model
+    (uniform ``num_experts`` × ``intermediate_size``).
+
+    The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
+    (pre-pruning) values for provenance; the width table is authoritative for
+    the built architecture.
+    """
+
+    model_type = "pruned_olmoe"
+
+    def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
+        super().__init__(**kwargs)
+        self.expert_widths = expert_widths

healed/mixceonly_keep50/vllm_live/generation_config.json

@@ -0,0 +1,6 @@
+{
+  "_from_model_config": true,
+  "eos_token_id": 50279,
+  "pad_token_id": 1,
+  "transformers_version": "4.57.6"
+}

healed/mixceonly_keep50/vllm_live/modeling_pruned_olmoe.py

@@ -0,0 +1,66 @@
+"""GLEAN-pruned OLMoE: HF-loadable model with ragged (variable-width) experts.
+
+Pattern follows hbfreed/variable-flex-olmo's PrunedFlexOlmoForCausalLM
+(docs/recon/prior-work-hbfreed.md), generalized from one scalar width to a
+per-(layer, expert) width table: ``super().__init__`` builds the uniform
+architecture from the config, then every MoE block is rebuilt to its pruned
+shape — surviving experts only, each at its own width, router sliced to
+match — so the state dict aligns exactly with what
+``glean.prune.prune_channels_global`` leaves behind.
+
+Caveat: ``output_router_logits=True`` (the load-balancing aux loss) assumes a
+uniform ``config.num_experts`` and is unsupported on ragged models.
+"""
+
+import torch.nn as nn
+from transformers.activations import ACT2FN
+from transformers.models.olmoe.modeling_olmoe import OlmoeForCausalLM
+
+from .configuration_pruned_olmoe import PrunedOlmoeConfig
+
+
+class RaggedOlmoeMLP(nn.Module):
+    """OlmoeMLP with an explicit intermediate width (SwiGLU, no biases)."""
+
+    def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
+        super().__init__()
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.gate_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.up_proj = nn.Linear(hidden_size, intermediate_size, bias=False)
+        self.down_proj = nn.Linear(intermediate_size, hidden_size, bias=False)
+        self.act_fn = ACT2FN[hidden_act]
+
+    def forward(self, x):
+        return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+
+
+class PrunedOlmoeForCausalLM(OlmoeForCausalLM):
+    """OLMoE with per-layer surviving-expert lists at per-expert widths."""
+
+    config_class = PrunedOlmoeConfig
+
+    def __init__(self, config: PrunedOlmoeConfig):
+        super().__init__(config)
+        widths_table = getattr(config, "expert_widths", None)
+        if widths_table is None:
+            return  # unpruned: plain OLMoE
+        if len(widths_table) != len(self.model.layers):
+            raise ValueError(
+                f"expert_widths has {len(widths_table)} rows but the model has "
+                f"{len(self.model.layers)} decoder layers"
+            )
+        for layer, widths in zip(self.model.layers, widths_table):
+            if any(w <= 0 for w in widths):
+                raise ValueError("expert_widths must list surviving experts only (>0)")
+            block = layer.mlp
+            if len(widths) < block.top_k:
+                raise ValueError(
+                    f"a layer keeps {len(widths)} experts < top_k={block.top_k}"
+                )
+            block.num_experts = len(widths)
+            block.gate = nn.Linear(config.hidden_size, len(widths), bias=False)
+            block.experts = nn.ModuleList(
+                RaggedOlmoeMLP(config.hidden_size, w, config.hidden_act)
+                for w in widths
+            )

healed/mixceonly_keep50/vllm_live/special_tokens_map.json

@@ -0,0 +1,23 @@
+{
+  "bos_token": {
+    "content": "|||IP_ADDRESS|||",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "eos_token": {
+    "content": "|||IP_ADDRESS|||",
+    "lstrip": false,
+    "normalized": true,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<pad>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

healed/mixceonly_keep50/vllm_live/tokenizer_config.json

@@ -0,0 +1,247 @@
+{
+  "add_bos_token": false,
+  "add_eos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|padding|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50254": {
+      "content": "                        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50255": {
+      "content": "                       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50256": {
+      "content": "                      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50257": {
+      "content": "                     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50258": {
+      "content": "                    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50259": {
+      "content": "                   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50260": {
+      "content": "                  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50261": {
+      "content": "                 ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50262": {
+      "content": "                ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50263": {
+      "content": "               ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50264": {
+      "content": "              ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50265": {
+      "content": "             ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50266": {
+      "content": "            ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50267": {
+      "content": "           ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50268": {
+      "content": "          ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50269": {
+      "content": "         ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50270": {
+      "content": "        ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50271": {
+      "content": "       ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50272": {
+      "content": "      ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50273": {
+      "content": "     ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50274": {
+      "content": "    ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50275": {
+      "content": "   ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50276": {
+      "content": "  ",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50277": {
+      "content": "|||EMAIL_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50278": {
+      "content": "|||PHONE_NUMBER|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "50279": {
+      "content": "|||IP_ADDRESS|||",
+      "lstrip": false,
+      "normalized": true,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "50280": {
+      "content": "<pad>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "bos_token": "|||IP_ADDRESS|||",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "|||IP_ADDRESS|||",
+  "extra_special_tokens": {},
+  "model_max_length": 1000000000000000019884624838656,
+  "pad_token": "<pad>",
+  "tokenizer_class": "GPTNeoXTokenizer",
+  "unk_token": null
+}

healed/mixceonly_keep50/wandb/debug-internal.log

@@ -0,0 +1,11 @@
+{"time":"2026-08-01T12:14:40.930763762-07:00","level":"INFO","msg":"wandb-core"}
+{"time":"2026-08-01T12:14:40.930810353-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-08-01T12:14:41.050876465-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T12:14:41.050913746-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T12:14:41.050952917-07:00","level":"INFO","msg":"stream: created new stream","id":"umzrxo81"}
+{"time":"2026-08-01T12:14:41.05106738-07:00","level":"INFO","msg":"handler: started"}
+{"time":"2026-08-01T12:14:41.051150532-07:00","level":"INFO","msg":"stream: started"}
+{"time":"2026-08-01T12:14:41.051193074-07:00","level":"INFO","msg":"writer: started","stream_id":"umzrxo81"}
+{"time":"2026-08-01T12:14:41.051238735-07:00","level":"INFO","msg":"sender: started"}
+{"time":"2026-08-01T12:14:41.065739707-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T12:14:41.065767307-07:00","level":"WARN","msg":"runupserter: server does not expand metric globs but the x_server_side_expand_glob_metrics setting is set; ignoring"}

healed/mixceonly_keep50/wandb/debug.log

@@ -0,0 +1,19 @@
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_setup.py:_flush():81] Configure stats pid to 13469
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug.log
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug-internal.log
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:init():768] calling init triggers
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'student': 'outputs/pruned/glean-0125inst-math-keep50', 'teacher': 'allenai/OLMoE-1B-7B-0125-Instruct', 'training_mode': 'on-policy', 'kl_direction': 'reverse', 'dataset': 'allenai/Dolci-Instruct-RL', 'dataset_sources': None, 'max_difficulty': None, 'trajectories': 'outputs/teacher_trajectories/dolci_math_curated.jsonl', 'trajectory_dataset': 'allenai/Dolci-Instruct-RL', 'off_policy_frames': 'chat', 'off_policy_max_seq_len': 2048, 'topk_targets': None, 'max_loss_tokens': None, 'loss_tokens_per_step': None, 'teacher_device': 'cuda:0', 'student_device': 'cuda:1', 'lr': 3e-05, 'optimizer': 'adamw8bit', 'weight_decay': 0.1, 'epochs': 2, 'prompts_per_step': 256, 'group_size': 4, 'rollout_batch': 64, 'micro_batch': 4, 'max_new_tokens': 2048, 'max_prompt_len': 1024, 'warmup_steps': 10, 'max_grad_norm': 1.0, 'eval_every': 10, 'gsm8k_every': 20, 'gsm8k_n': 256, 'gsm8k_batch': 16, 'gsm8k_max_new_tokens': 1024, 'gsm8k_frames': 'chat', 'save_every': 1000, 'out_dir': 'outputs/healed/mixceonly_keep50', 'sweep': 60, 'wandb': True, 'wandb_project': 'glean-heal', 'wandb_run_name': 'mixce-keep50-s1223', 'wandb_run_id': None, 'wandb_resume': None, 'wandb_mode': 'offline', 'no_wandb_sync': False, 'debug': False, 'resume_from': None, 'start_step': 0, 'no_grad_checkpointing': False, 'seed': 1223, 'no_teacher_overlap': False, 'sync_checkpoints': False, 'rollout_engine': 'vllm', 'vllm_gpu': '2', 'vllm_port': 8377, 'vllm_refresh_every': 1, 'vllm_serve_bin': 'vllm-plugin/.venv/bin/python', 'vllm_gpu_mem_util': 0.85, 'liger_loss': True, 'gold_mix_lambda': 0.5, 'gold_topk_targets': 'outputs/teacher_trajectories/dolci_combined_top128', 'gold_loss': 'ce', 'gold_mix_decay': 0.0, 'fast_teacher': True, 'reference_kl_beta': 0.0, 'drop_truncated_rollouts': False, 'vllm_max_model_len': None, 'vllm_refresh_mode': 'reload', 'vllm_live_dir': None, 'resolved_kl_direction': 'reverse', '_wandb': {}}
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:init():816] starting backend
+2026-08-01 12:14:40,921 INFO    MainThread:13469 [wandb_init.py:init():831] sending inform_init request
+2026-08-01 12:14:41,052 INFO    MainThread:13469 [wandb_init.py:init():836] backend started and connected
+2026-08-01 12:14:41,054 INFO    MainThread:13469 [wandb_init.py:init():906] updated telemetry
+2026-08-01 12:14:41,062 INFO    MainThread:13469 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-08-01 12:14:41,068 INFO    MainThread:13469 [wandb_init.py:init():974] starting run threads in backend
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_console_start():2523] atexit reg
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-08-01 12:14:41,199 INFO    MainThread:13469 [wandb_init.py:init():1012] run started, returning control to user process

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/files/config.yaml

@@ -0,0 +1,353 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    python_version: 3.12.12
+    cli_version: 0.28.0
+    framework: huggingface
+    huggingface_version: 4.57.6
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    start_time: 1785610770
+    t:
+      1:
+      - 1
+      - 5
+      - 11
+      - 49
+      - 51
+      - 53
+      - 71
+      2:
+      - 1
+      - 5
+      - 11
+      - 49
+      - 51
+      - 53
+      - 71
+      3:
+      - 2
+      - 4
+      - 7
+      - 13
+      - 16
+      - 42
+      4: 3.12.12
+      5: 0.28.0
+      6: 4.57.6
+      13: linux-x86_64
+    e:
+      iem7c11crq1tx9zum49b4a6wk3d9vokq:
+        os: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35
+        python: CPython 3.12.12
+        started_at: '2026-08-01T18:59:30.135060Z'
+        args:
+        - --student
+        - outputs/pruned/glean-0125inst-math-keep50
+        - --teacher
+        - allenai/OLMoE-1B-7B-0125-Instruct
+        - --training-mode
+        - on-policy
+        - --kl-direction
+        - reverse
+        - --dataset
+        - allenai/Dolci-Instruct-RL
+        - --lr
+        - 3e-5
+        - --optimizer
+        - adamw8bit
+        - --epochs
+        - '2'
+        - --sweep
+        - '60'
+        - --max-new-tokens
+        - '2048'
+        - --group-size
+        - '4'
+        - --prompts-per-step
+        - '256'
+        - --micro-batch
+        - '4'
+        - --seed
+        - '1223'
+        - --save-every
+        - '1000'
+        - --gsm8k-every
+        - '20'
+        - --gsm8k-max-new-tokens
+        - '1024'
+        - --fast-teacher
+        - --liger-loss
+        - --gold-mix-lambda
+        - '0.5'
+        - --gold-loss
+        - ce
+        - --gold-topk-targets
+        - outputs/teacher_trajectories/dolci_combined_top128
+        - --rollout-engine
+        - vllm
+        - --vllm-gpu
+        - '2'
+        - --vllm-refresh-every
+        - '1'
+        - --student-device
+        - cuda:1
+        - --teacher-device
+        - cuda:0
+        - --out-dir
+        - outputs/healed/mixceonly_keep50
+        - --wandb
+        - --wandb-project
+        - glean-heal
+        - --wandb-run-name
+        - mixce-keep50-s1223
+        program: /home/henry/Documents/PythonProjects/variable-reap/scripts/11_distill_on_policy.py
+        code_path: scripts/11_distill_on_policy.py
+        code_path_local: scripts/11_distill_on_policy.py
+        git:
+          remote_url: https://github.com/hbfreed/variable-reap.git
+          commit: 22365a35ebf321b7ccfa3430a715a153e787d3ee
+        root: outputs/healed/mixceonly_keep50
+        host: pop-os
+        executable: /home/henry/Documents/PythonProjects/variable-reap/.venv/bin/python3
+        cpu_count: 24
+        cpu_count_logical: 48
+        gpu_type: NVIDIA GeForce RTX 3090
+        gpu_count: 3
+        disk:
+          /:
+            total: '1958315118592'
+            used: '1293342965760'
+        memory:
+          total: '134898364416'
+        gpu_nvidia:
+        - name: NVIDIA GeForce RTX 3090
+          memory_total: '25769803776'
+          cuda_cores: 10496
+          architecture: Ampere
+          uuid: GPU-8ca70870-ddf2-d274-bcc5-182c2075bced
+        - name: NVIDIA GeForce RTX 3090
+          memory_total: '25769803776'
+          cuda_cores: 10496
+          architecture: Ampere
+          uuid: GPU-a6acf07f-31f5-618f-a5d0-c0017e7e2e27
+        - name: NVIDIA GeForce RTX 3090
+          memory_total: '25769803776'
+          cuda_cores: 10496
+          architecture: Ampere
+          uuid: GPU-864c54df-0130-7780-e271-8a5551d1733f
+        cuda_version: '13.0'
+        writer_id: iem7c11crq1tx9zum49b4a6wk3d9vokq
+    m:
+    - 1: step
+      6:
+      - 3
+student:
+  desc: null
+  value: outputs/pruned/glean-0125inst-math-keep50
+teacher:
+  desc: null
+  value: allenai/OLMoE-1B-7B-0125-Instruct
+training_mode:
+  desc: null
+  value: on-policy
+kl_direction:
+  desc: null
+  value: reverse
+dataset:
+  desc: null
+  value: allenai/Dolci-Instruct-RL
+dataset_sources:
+  desc: null
+  value: null
+max_difficulty:
+  desc: null
+  value: null
+trajectories:
+  desc: null
+  value: outputs/teacher_trajectories/dolci_math_curated.jsonl
+trajectory_dataset:
+  desc: null
+  value: allenai/Dolci-Instruct-RL
+off_policy_frames:
+  desc: null
+  value: chat
+off_policy_max_seq_len:
+  desc: null
+  value: 2048
+topk_targets:
+  desc: null
+  value: null
+max_loss_tokens:
+  desc: null
+  value: null
+loss_tokens_per_step:
+  desc: null
+  value: null
+teacher_device:
+  desc: null
+  value: cuda:0
+student_device:
+  desc: null
+  value: cuda:1
+lr:
+  desc: null
+  value: 3.0e-05
+optimizer:
+  desc: null
+  value: adamw8bit
+weight_decay:
+  desc: null
+  value: 0.1
+epochs:
+  desc: null
+  value: 2
+prompts_per_step:
+  desc: null
+  value: 256
+group_size:
+  desc: null
+  value: 4
+rollout_batch:
+  desc: null
+  value: 64
+micro_batch:
+  desc: null
+  value: 4
+max_new_tokens:
+  desc: null
+  value: 2048
+max_prompt_len:
+  desc: null
+  value: 1024
+warmup_steps:
+  desc: null
+  value: 10
+max_grad_norm:
+  desc: null
+  value: 1.0
+eval_every:
+  desc: null
+  value: 10
+gsm8k_every:
+  desc: null
+  value: 20
+gsm8k_n:
+  desc: null
+  value: 256
+gsm8k_batch:
+  desc: null
+  value: 16
+gsm8k_max_new_tokens:
+  desc: null
+  value: 1024
+gsm8k_frames:
+  desc: null
+  value: chat
+save_every:
+  desc: null
+  value: 1000
+out_dir:
+  desc: null
+  value: outputs/healed/mixceonly_keep50
+sweep:
+  desc: null
+  value: 60
+wandb:
+  desc: null
+  value: true
+wandb_project:
+  desc: null
+  value: glean-heal
+wandb_run_name:
+  desc: null
+  value: mixce-keep50-s1223
+wandb_run_id:
+  desc: null
+  value: null
+wandb_resume:
+  desc: null
+  value: null
+wandb_mode:
+  desc: null
+  value: offline
+no_wandb_sync:
+  desc: null
+  value: false
+debug:
+  desc: null
+  value: false
+resume_from:
+  desc: null
+  value: null
+start_step:
+  desc: null
+  value: 0
+no_grad_checkpointing:
+  desc: null
+  value: false
+seed:
+  desc: null
+  value: 1223
+no_teacher_overlap:
+  desc: null
+  value: false
+sync_checkpoints:
+  desc: null
+  value: false
+rollout_engine:
+  desc: null
+  value: vllm
+vllm_gpu:
+  desc: null
+  value: '2'
+vllm_port:
+  desc: null
+  value: 8377
+vllm_refresh_every:
+  desc: null
+  value: 1
+vllm_serve_bin:
+  desc: null
+  value: vllm-plugin/.venv/bin/python
+vllm_gpu_mem_util:
+  desc: null
+  value: 0.85
+liger_loss:
+  desc: null
+  value: true
+gold_mix_lambda:
+  desc: null
+  value: 0.5
+gold_topk_targets:
+  desc: null
+  value: outputs/teacher_trajectories/dolci_combined_top128
+gold_loss:
+  desc: null
+  value: ce
+gold_mix_decay:
+  desc: null
+  value: 0.0
+fast_teacher:
+  desc: null
+  value: true
+reference_kl_beta:
+  desc: null
+  value: 0.0
+drop_truncated_rollouts:
+  desc: null
+  value: false
+vllm_max_model_len:
+  desc: null
+  value: null
+vllm_refresh_mode:
+  desc: null
+  value: reload
+vllm_live_dir:
+  desc: null
+  value: null
+resolved_kl_direction:
+  desc: null
+  value: reverse

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/files/output.log

@@ -0,0 +1,8 @@
+Loading checkpoint shards: 100%|██████████| 3/3 [00:05<00:00,  1.76s/it]
+Loading checkpoint shards: 100%|██████████| 2/2 [00:03<00:00,  1.53s/it]
+Filter: 100%|██████████| 169964/169964 [00:10<00:00, 16958.55 examples/s]
+teacher converted to fused MoE path (allenai/OLMoE-1B-7B-0125-Instruct)
+starting vllm rollout server on GPU 2 (port 8377) ...
+vllm server healthy in 46s
+mixture distillation: 58360 gold trajectories from outputs/teacher_trajectories/dolci_combined_top128, lambda=0.5 decay=0.0
+168191 prompts | 2627 steps/epoch | 60 total steps | student params 3.70B | teacher overlap=True

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/files/requirements.txt

@@ -0,0 +1,130 @@
+certifi==2026.6.17
+fonttools==4.63.0
+urllib3==2.7.0
+requests==2.34.2
+nvidia-cublas-cu12==12.6.4.1
+packaging==26.2
+regex==2026.6.28
+portalocker==3.2.0
+safetensors==0.8.0
+datasets==5.0.0
+sqlitedict==2.1.0
+narwhals==2.23.0
+latex2sympy2_extended==1.11.0
+smmap==5.0.3
+colorama==0.4.6
+anyio==4.14.1
+cycler==0.12.1
+pytablewriter==1.2.1
+pydantic_core==2.46.4
+huggingface_hub==0.36.2
+charset-normalizer==3.4.8
+nvidia-cusolver-cu12==11.7.1.2
+joblib==1.5.3
+typing-inspection==0.4.2
+wandb==0.28.0
+bitsandbytes==0.49.2
+kiwisolver==1.5.0
+cuda-toolkit==12.6.3
+propcache==0.5.2
+Jinja2==3.1.6
+cuda-bindings==12.9.7
+pluggy==1.6.0
+pytest==9.1.1
+fsspec==2026.4.0
+tqdm==4.68.3
+psutil==7.2.2
+pydantic==2.13.4
+typepy==1.3.5
+xxhash==3.8.1
+DataProperty==1.1.1
+aiosignal==1.4.0
+annotated-types==0.7.0
+Pygments==2.20.0
+zstandard==0.25.0
+threadpoolctl==3.6.0
+liger_kernel==0.8.1
+nvidia-cudnn-cu12==9.10.2.21
+evaluate==0.4.6
+pillow==12.3.0
+GitPython==3.1.50
+httpx==0.28.1
+lm_eval==0.4.12
+pytz==2026.2
+mbstrdecoder==1.1.5
+immutabledict==4.3.1
+dill==0.4.1
+mpmath==1.3.0
+yarl==1.24.2
+word2number==1.1
+networkx==3.6.1
+einops==0.8.2
+tabulate==0.10.0
+pathvalidate==3.3.1
+tenacity==9.1.4
+sentry-sdk==2.64.0
+nvidia-nvshmem-cu12==3.4.5
+hf-xet==1.5.1
+six==1.17.0
+pyarrow==24.0.0
+nvidia-curand-cu12==10.3.7.77
+cuda-pathfinder==1.5.6
+sacrebleu==2.6.0
+triton==3.7.1
+frozenlist==1.8.0
+antlr4-python3-runtime==4.11.0
+tabledata==1.3.5
+h11==0.16.0
+nvidia-cusparselt-cu12==0.7.1
+nvidia-nvjitlink-cu12==12.6.85
+scipy==1.18.0
+rouge_score==0.1.2
+transformers==4.57.6
+multiprocess==0.70.19
+absl-py==2.5.0
+accelerate==1.14.0
+click==8.4.2
+lxml==6.1.1
+defusedxml==0.7.1
+setuptools==81.0.0
+nvidia-cuda-cupti-cu12==12.6.80
+httpcore==1.0.9
+langdetect==1.0.9
+nltk==3.10.0
+iniconfig==2.3.0
+nvidia-cuda-runtime-cu12==12.6.77
+nvidia-cusparse-cu12==12.5.4.2
+matplotlib==3.11.0
+protobuf==7.35.1
+nvidia-cufft-cu12==11.3.0.4
+pyparsing==3.3.2
+chardet==6.0.0.post1
+typing_extensions==4.16.0
+stanford-stk==0.7.1
+aiohappyeyeballs==2.7.1
+MarkupSafe==3.0.3
+nvidia-cufile-cu12==1.11.1.6
+tcolorpy==0.1.7
+scikit-learn==1.9.0
+tokenizers==0.22.2
+attrs==26.1.0
+gitdb==4.0.12
+aiohttp==3.14.1
+glean==0.0.1
+platformdirs==4.10.0
+megablocks==0.11.0.dev0
+nvidia-cuda-nvrtc-cu12==12.6.85
+pandas==3.0.3
+multidict==6.7.1
+torch==2.12.1+cu126
+contourpy==1.3.3
+nvidia-nvtx-cu12==12.6.77
+math-verify==0.9.0
+filelock==3.29.5
+idna==3.18
+nvidia-nccl-cu12==2.29.3
+python-dateutil==2.9.0.post0
+more-itertools==11.1.0
+numpy==2.0.2
+sympy==1.14.0
+PyYAML==6.0.3

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/files/wandb-metadata.json

@@ -0,0 +1 @@
+{"os": "Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35", "python": "CPython 3.12.12", "started_at": "2026-08-01T18:59:30.135060Z", "args": ["--student", "outputs/pruned/glean-0125inst-math-keep50", "--teacher", "allenai/OLMoE-1B-7B-0125-Instruct", "--training-mode", "on-policy", "--kl-direction", "reverse", "--dataset", "allenai/Dolci-Instruct-RL", "--lr", "3e-5", "--optimizer", "adamw8bit", "--epochs", "2", "--sweep", "60", "--max-new-tokens", "2048", "--group-size", "4", "--prompts-per-step", "256", "--micro-batch", "4", "--seed", "1223", "--save-every", "1000", "--gsm8k-every", "20", "--gsm8k-max-new-tokens", "1024", "--fast-teacher", "--liger-loss", "--gold-mix-lambda", "0.5", "--gold-loss", "ce", "--gold-topk-targets", "outputs/teacher_trajectories/dolci_combined_top128", "--rollout-engine", "vllm", "--vllm-gpu", "2", "--vllm-refresh-every", "1", "--student-device", "cuda:1", "--teacher-device", "cuda:0", "--out-dir", "outputs/healed/mixceonly_keep50", "--wandb", "--wandb-project", "glean-heal", "--wandb-run-name", "mixce-keep50-s1223"], "program": "/home/henry/Documents/PythonProjects/variable-reap/scripts/11_distill_on_policy.py", "code_path": "scripts/11_distill_on_policy.py", "code_path_local": "scripts/11_distill_on_policy.py", "git": {"remote_url": "https://github.com/hbfreed/variable-reap.git", "commit": "22365a35ebf321b7ccfa3430a715a153e787d3ee"}, "root": "outputs/healed/mixceonly_keep50", "host": "pop-os", "executable": "/home/henry/Documents/PythonProjects/variable-reap/.venv/bin/python3", "cpu_count": 24, "cpu_count_logical": 48, "gpu_type": "NVIDIA GeForce RTX 3090", "gpu_count": 3, "disk": {"/": {"total": "1958315118592", "used": "1293342965760"}}, "memory": {"total": "134898364416"}, "gpu_nvidia": [{"name": "NVIDIA GeForce RTX 3090", "memory_total": "25769803776", "cuda_cores": 10496, "architecture": "Ampere", "uuid": "GPU-8ca70870-ddf2-d274-bcc5-182c2075bced"}, {"name": "NVIDIA GeForce RTX 3090", "memory_total": "25769803776", "cuda_cores": 10496, "architecture": "Ampere", "uuid": "GPU-a6acf07f-31f5-618f-a5d0-c0017e7e2e27"}, {"name": "NVIDIA GeForce RTX 3090", "memory_total": "25769803776", "cuda_cores": 10496, "architecture": "Ampere", "uuid": "GPU-864c54df-0130-7780-e271-8a5551d1733f"}], "cuda_version": "13.0", "writer_id": "iem7c11crq1tx9zum49b4a6wk3d9vokq"}

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_runtime": 252}

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/logs/debug-core.log

@@ -0,0 +1,17 @@
+{"time":"2026-08-01T11:59:30.401252082-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpoi5y04qe/port-10354.txt","pid":10354,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-08-01T11:59:30.4022684-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":10354}
+{"time":"2026-08-01T11:59:30.402180828-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-10354-10470-553243608/socket","Net":"unix"}}
+{"time":"2026-08-01T11:59:30.573538202-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-08-01T11:59:30.581774192-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"009xr5f3","id":"1(@)"}
+{"time":"2026-08-01T11:59:30.810763703-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"009xr5f3","id":"1(@)"}
+{"time":"2026-08-01T12:03:42.981109151-07:00","level":"INFO","msg":"handleInformFinish: finish message received","streamId":"009xr5f3","id":"1(@)"}
+{"time":"2026-08-01T12:03:42.986634529-07:00","level":"INFO","msg":"handleInformFinish: stream closed","streamId":"009xr5f3","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.755813028-07:00","level":"INFO","msg":"handleInformTeardown: server teardown initiated","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.75586421-07:00","level":"INFO","msg":"handleInformTeardown: server shutdown complete","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.75588066-07:00","level":"INFO","msg":"connection: closing","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.755907351-07:00","level":"INFO","msg":"server is shutting down"}
+{"time":"2026-08-01T12:03:43.755943942-07:00","level":"INFO","msg":"connection: closed successfully","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.755958353-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.756010244-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"1(@)"}
+{"time":"2026-08-01T12:03:43.756087286-07:00","level":"INFO","msg":"server: listener closed","addr":{"Name":"/tmp/wandb-10354-10470-553243608/socket","Net":"unix"}}
+{"time":"2026-08-01T12:03:43.756147248-07:00","level":"INFO","msg":"server is closed"}

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/logs/debug-internal.log

@@ -0,0 +1,16 @@
+{"time":"2026-08-01T11:59:30.581923106-07:00","level":"INFO","msg":"wandb-core"}
+{"time":"2026-08-01T11:59:30.581954627-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-08-01T11:59:30.808966253-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T11:59:30.808997723-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T11:59:30.809838707-07:00","level":"INFO","msg":"stream: created new stream","id":"009xr5f3"}
+{"time":"2026-08-01T11:59:30.810596098-07:00","level":"INFO","msg":"handler: started"}
+{"time":"2026-08-01T11:59:30.810747402-07:00","level":"INFO","msg":"stream: started"}
+{"time":"2026-08-01T11:59:30.811480903-07:00","level":"INFO","msg":"writer: started","stream_id":"009xr5f3"}
+{"time":"2026-08-01T11:59:30.811518994-07:00","level":"INFO","msg":"sender: started"}
+{"time":"2026-08-01T11:59:30.82752934-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T11:59:30.827577921-07:00","level":"WARN","msg":"runupserter: server does not expand metric globs but the x_server_side_expand_glob_metrics setting is set; ignoring"}
+{"time":"2026-08-01T12:03:42.977586459-07:00","level":"INFO","msg":"handler: operation stats","stats":{}}
+{"time":"2026-08-01T12:03:42.981134481-07:00","level":"INFO","msg":"stream: finishing up"}
+{"time":"2026-08-01T12:03:42.981145452-07:00","level":"INFO","msg":"handler: closed"}
+{"time":"2026-08-01T12:03:42.981194183-07:00","level":"INFO","msg":"sender: closed"}
+{"time":"2026-08-01T12:03:42.981204133-07:00","level":"INFO","msg":"stream: all finished"}

healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/logs/debug.log

@@ -0,0 +1,23 @@
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_setup.py:_flush():81] Configure stats pid to 10354
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/logs/debug.log
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/healed/mixceonly_keep50/wandb/offline-run-20260801_115930-009xr5f3/logs/debug-internal.log
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_init.py:init():768] calling init triggers
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'student': 'outputs/pruned/glean-0125inst-math-keep50', 'teacher': 'allenai/OLMoE-1B-7B-0125-Instruct', 'training_mode': 'on-policy', 'kl_direction': 'reverse', 'dataset': 'allenai/Dolci-Instruct-RL', 'dataset_sources': None, 'max_difficulty': None, 'trajectories': 'outputs/teacher_trajectories/dolci_math_curated.jsonl', 'trajectory_dataset': 'allenai/Dolci-Instruct-RL', 'off_policy_frames': 'chat', 'off_policy_max_seq_len': 2048, 'topk_targets': None, 'max_loss_tokens': None, 'loss_tokens_per_step': None, 'teacher_device': 'cuda:0', 'student_device': 'cuda:1', 'lr': 3e-05, 'optimizer': 'adamw8bit', 'weight_decay': 0.1, 'epochs': 2, 'prompts_per_step': 256, 'group_size': 4, 'rollout_batch': 64, 'micro_batch': 4, 'max_new_tokens': 2048, 'max_prompt_len': 1024, 'warmup_steps': 10, 'max_grad_norm': 1.0, 'eval_every': 10, 'gsm8k_every': 20, 'gsm8k_n': 256, 'gsm8k_batch': 16, 'gsm8k_max_new_tokens': 1024, 'gsm8k_frames': 'chat', 'save_every': 1000, 'out_dir': 'outputs/healed/mixceonly_keep50', 'sweep': 60, 'wandb': True, 'wandb_project': 'glean-heal', 'wandb_run_name': 'mixce-keep50-s1223', 'wandb_run_id': None, 'wandb_resume': None, 'wandb_mode': 'offline', 'no_wandb_sync': False, 'debug': False, 'resume_from': None, 'start_step': 0, 'no_grad_checkpointing': False, 'seed': 1223, 'no_teacher_overlap': False, 'sync_checkpoints': False, 'rollout_engine': 'vllm', 'vllm_gpu': '2', 'vllm_port': 8377, 'vllm_refresh_every': 1, 'vllm_serve_bin': 'vllm-plugin/.venv/bin/python', 'vllm_gpu_mem_util': 0.85, 'liger_loss': True, 'gold_mix_lambda': 0.5, 'gold_topk_targets': 'outputs/teacher_trajectories/dolci_combined_top128', 'gold_loss': 'ce', 'gold_mix_decay': 0.0, 'fast_teacher': True, 'reference_kl_beta': 0.0, 'drop_truncated_rollouts': False, 'vllm_max_model_len': None, 'vllm_refresh_mode': 'reload', 'vllm_live_dir': None, 'resolved_kl_direction': 'reverse', '_wandb': {}}
+2026-08-01 11:59:30,136 INFO    MainThread:10354 [wandb_init.py:init():816] starting backend
+2026-08-01 11:59:30,573 INFO    MainThread:10354 [wandb_init.py:init():831] sending inform_init request
+2026-08-01 11:59:30,811 INFO    MainThread:10354 [wandb_init.py:init():836] backend started and connected
+2026-08-01 11:59:30,814 INFO    MainThread:10354 [wandb_init.py:init():906] updated telemetry
+2026-08-01 11:59:30,822 INFO    MainThread:10354 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-08-01 11:59:30,831 INFO    MainThread:10354 [wandb_init.py:init():974] starting run threads in backend
+2026-08-01 11:59:31,001 INFO    MainThread:10354 [wandb_run.py:_console_start():2523] atexit reg
+2026-08-01 11:59:31,001 INFO    MainThread:10354 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-08-01 11:59:31,001 INFO    MainThread:10354 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-08-01 11:59:31,001 INFO    MainThread:10354 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-08-01 11:59:31,002 INFO    MainThread:10354 [wandb_init.py:init():1012] run started, returning control to user process
+2026-08-01 12:03:42,974 INFO    MainThread:10354 [wandb_run.py:_finish():2285] finishing run glean-heal/009xr5f3
+2026-08-01 12:03:42,975 INFO    MainThread:10354 [wandb_run.py:_atexit_cleanup():2490] got exitcode: 0
+2026-08-01 12:03:42,975 INFO    MainThread:10354 [wandb_run.py:_restore():2472] restore
+2026-08-01 12:03:42,975 INFO    MainThread:10354 [wandb_run.py:_restore():2478] restore done

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/files/config.yaml

@@ -0,0 +1,432 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    python_version: 3.12.12
+    cli_version: 0.28.0
+    framework: huggingface
+    huggingface_version: 4.57.6
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    start_time: 1785611681
+    t:
+      1:
+      - 1
+      - 5
+      - 11
+      - 49
+      - 51
+      - 53
+      - 71
+      2:
+      - 1
+      - 5
+      - 11
+      - 49
+      - 51
+      - 53
+      - 71
+      3:
+      - 4
+      - 7
+      - 13
+      - 16
+      - 42
+      4: 3.12.12
+      5: 0.28.0
+      6: 4.57.6
+      13: linux-x86_64
+    e:
+      cdl6ysspj1b4fap16js7r9iidg0igxex:
+        os: Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35
+        python: CPython 3.12.12
+        started_at: '2026-08-01T19:14:40.480723Z'
+        args:
+        - --student
+        - outputs/pruned/glean-0125inst-math-keep50
+        - --teacher
+        - allenai/OLMoE-1B-7B-0125-Instruct
+        - --training-mode
+        - on-policy
+        - --kl-direction
+        - reverse
+        - --dataset
+        - allenai/Dolci-Instruct-RL
+        - --lr
+        - 3e-5
+        - --optimizer
+        - adamw8bit
+        - --epochs
+        - '2'
+        - --sweep
+        - '60'
+        - --max-new-tokens
+        - '2048'
+        - --group-size
+        - '4'
+        - --prompts-per-step
+        - '256'
+        - --micro-batch
+        - '4'
+        - --seed
+        - '1223'
+        - --save-every
+        - '1000'
+        - --gsm8k-every
+        - '20'
+        - --gsm8k-max-new-tokens
+        - '1024'
+        - --fast-teacher
+        - --liger-loss
+        - --gold-mix-lambda
+        - '0.5'
+        - --gold-loss
+        - ce
+        - --gold-topk-targets
+        - outputs/teacher_trajectories/dolci_combined_top128
+        - --rollout-engine
+        - vllm
+        - --vllm-gpu
+        - '2'
+        - --vllm-refresh-every
+        - '1'
+        - --student-device
+        - cuda:1
+        - --teacher-device
+        - cuda:0
+        - --out-dir
+        - outputs/healed/mixceonly_keep50
+        - --wandb
+        - --wandb-project
+        - glean-heal
+        - --wandb-run-name
+        - mixce-keep50-s1223
+        program: /home/henry/Documents/PythonProjects/variable-reap/scripts/11_distill_on_policy.py
+        code_path: scripts/11_distill_on_policy.py
+        code_path_local: scripts/11_distill_on_policy.py
+        git:
+          remote_url: https://github.com/hbfreed/variable-reap.git
+          commit: 22365a35ebf321b7ccfa3430a715a153e787d3ee
+        root: outputs/healed/mixceonly_keep50
+        host: pop-os
+        executable: /home/henry/Documents/PythonProjects/variable-reap/.venv/bin/python
+        cpu_count: 24
+        cpu_count_logical: 48
+        gpu_type: NVIDIA GeForce RTX 3090
+        gpu_count: 3
+        disk:
+          /:
+            total: '1958315118592'
+            used: '1293356351488'
+        memory:
+          total: '134898364416'
+        gpu_nvidia:
+        - name: NVIDIA GeForce RTX 3090
+          memory_total: '25769803776'
+          cuda_cores: 10496
+          architecture: Ampere
+          uuid: GPU-8ca70870-ddf2-d274-bcc5-182c2075bced
+        - name: NVIDIA GeForce RTX 3090
+          memory_total: '25769803776'
+          cuda_cores: 10496
+          architecture: Ampere
+          uuid: GPU-a6acf07f-31f5-618f-a5d0-c0017e7e2e27
+        - name: NVIDIA GeForce RTX 3090
+          memory_total: '25769803776'
+          cuda_cores: 10496
+          architecture: Ampere
+          uuid: GPU-864c54df-0130-7780-e271-8a5551d1733f
+        cuda_version: '13.0'
+        writer_id: cdl6ysspj1b4fap16js7r9iidg0igxex
+    m:
+    - 1: step
+      6:
+      - 3
+    - 1: reverse_kl
+      5: 1
+      6:
+      - 1
+    - 1: rep_ratio
+      5: 1
+      6:
+      - 1
+    - 1: mem_gb_teacher
+      5: 1
+      6:
+      - 1
+    - 1: cumulative_loss_tokens
+      5: 1
+      6:
+      - 1
+    - 1: grad_norm
+      5: 1
+      6:
+      - 1
+    - 1: lr
+      5: 1
+      6:
+      - 1
+    - 1: gold_lambda
+      5: 1
+      6:
+      - 1
+    - 1: t_data_s
+      5: 1
+      6:
+      - 1
+    - 1: t_rollout_s
+      5: 1
+      6:
+      - 1
+    - 1: epoch
+      5: 1
+      6:
+      - 1
+    - 1: training_mode
+      5: 1
+      6:
+      - 1
+    - 1: tokens
+      5: 1
+      6:
+      - 1
+    - 1: comp_len
+      5: 1
+      6:
+      - 1
+    - 1: t_refresh_s
+      5: 1
+      6:
+      - 1
+    - 1: finish_rate
+      5: 1
+      6:
+      - 1
+    - 1: dropped_truncated
+      5: 1
+      6:
+      - 1
+    - 1: gold_loss
+      5: 1
+      6:
+      - 1
+    - 1: t_step_s
+      5: 1
+      6:
+      - 1
+    - 1: mem_gb
+      5: 1
+      6:
+      - 1
+    - 1: gsm8k_quick_chat
+      5: 1
+      6:
+      - 1
+student:
+  desc: null
+  value: outputs/pruned/glean-0125inst-math-keep50
+teacher:
+  desc: null
+  value: allenai/OLMoE-1B-7B-0125-Instruct
+training_mode:
+  desc: null
+  value: on-policy
+kl_direction:
+  desc: null
+  value: reverse
+dataset:
+  desc: null
+  value: allenai/Dolci-Instruct-RL
+dataset_sources:
+  desc: null
+  value: null
+max_difficulty:
+  desc: null
+  value: null
+trajectories:
+  desc: null
+  value: outputs/teacher_trajectories/dolci_math_curated.jsonl
+trajectory_dataset:
+  desc: null
+  value: allenai/Dolci-Instruct-RL
+off_policy_frames:
+  desc: null
+  value: chat
+off_policy_max_seq_len:
+  desc: null
+  value: 2048
+topk_targets:
+  desc: null
+  value: null
+max_loss_tokens:
+  desc: null
+  value: null
+loss_tokens_per_step:
+  desc: null
+  value: null
+teacher_device:
+  desc: null
+  value: cuda:0
+student_device:
+  desc: null
+  value: cuda:1
+lr:
+  desc: null
+  value: 3.0e-05
+optimizer:
+  desc: null
+  value: adamw8bit
+weight_decay:
+  desc: null
+  value: 0.1
+epochs:
+  desc: null
+  value: 2
+prompts_per_step:
+  desc: null
+  value: 256
+group_size:
+  desc: null
+  value: 4
+rollout_batch:
+  desc: null
+  value: 64
+micro_batch:
+  desc: null
+  value: 4
+max_new_tokens:
+  desc: null
+  value: 2048
+max_prompt_len:
+  desc: null
+  value: 1024
+warmup_steps:
+  desc: null
+  value: 10
+max_grad_norm:
+  desc: null
+  value: 1.0
+eval_every:
+  desc: null
+  value: 10
+gsm8k_every:
+  desc: null
+  value: 20
+gsm8k_n:
+  desc: null
+  value: 256
+gsm8k_batch:
+  desc: null
+  value: 16
+gsm8k_max_new_tokens:
+  desc: null
+  value: 1024
+gsm8k_frames:
+  desc: null
+  value: chat
+save_every:
+  desc: null
+  value: 1000
+out_dir:
+  desc: null
+  value: outputs/healed/mixceonly_keep50
+sweep:
+  desc: null
+  value: 60
+wandb:
+  desc: null
+  value: true
+wandb_project:
+  desc: null
+  value: glean-heal
+wandb_run_name:
+  desc: null
+  value: mixce-keep50-s1223
+wandb_run_id:
+  desc: null
+  value: null
+wandb_resume:
+  desc: null
+  value: null
+wandb_mode:
+  desc: null
+  value: offline
+no_wandb_sync:
+  desc: null
+  value: false
+debug:
+  desc: null
+  value: false
+resume_from:
+  desc: null
+  value: null
+start_step:
+  desc: null
+  value: 0
+no_grad_checkpointing:
+  desc: null
+  value: false
+seed:
+  desc: null
+  value: 1223
+no_teacher_overlap:
+  desc: null
+  value: false
+sync_checkpoints:
+  desc: null
+  value: false
+rollout_engine:
+  desc: null
+  value: vllm
+vllm_gpu:
+  desc: null
+  value: '2'
+vllm_port:
+  desc: null
+  value: 8377
+vllm_refresh_every:
+  desc: null
+  value: 1
+vllm_serve_bin:
+  desc: null
+  value: vllm-plugin/.venv/bin/python
+vllm_gpu_mem_util:
+  desc: null
+  value: 0.85
+liger_loss:
+  desc: null
+  value: true
+gold_mix_lambda:
+  desc: null
+  value: 0.5
+gold_topk_targets:
+  desc: null
+  value: outputs/teacher_trajectories/dolci_combined_top128
+gold_loss:
+  desc: null
+  value: ce
+gold_mix_decay:
+  desc: null
+  value: 0.0
+fast_teacher:
+  desc: null
+  value: true
+reference_kl_beta:
+  desc: null
+  value: 0.0
+drop_truncated_rollouts:
+  desc: null
+  value: false
+vllm_max_model_len:
+  desc: null
+  value: null
+vllm_refresh_mode:
+  desc: null
+  value: reload
+vllm_live_dir:
+  desc: null
+  value: null
+resolved_kl_direction:
+  desc: null
+  value: reverse

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/files/output.log

@@ -0,0 +1,61 @@
+Loading checkpoint shards: 100%|██████████| 3/3 [00:00<00:00,  9.71it/s]
+Loading checkpoint shards: 100%|██████████| 2/2 [00:00<00:00,  7.89it/s]
+teacher converted to fused MoE path (allenai/OLMoE-1B-7B-0125-Instruct)
+starting vllm rollout server on GPU 2 (port 8377) ...
+vllm server healthy in 36s
+mixture distillation: 58360 gold trajectories from outputs/teacher_trajectories/dolci_combined_top128, lambda=0.5 decay=0.0
+168191 prompts | 2627 steps/epoch | 60 total steps | student params 3.70B | teacher overlap=True
+{"step": 1, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6070861224798003, "tokens": 174656, "cumulative_loss_tokens": 174656, "grad_norm": 4.1875, "lr": 6e-06, "finish_rate": 0.98, "comp_len": 682.2, "dropped_truncated": 0, "gold_loss": 0.4598, "gold_lambda": 0.5, "rep_ratio": 2.51, "t_data_s": 0.0, "t_rollout_s": 52.1, "t_step_s": 127.5, "t_refresh_s": 0.3, "mem_gb": 10.13, "mem_gb_teacher": 13.64}
+[eval step 1] sample: 'Misy fitsapana manokana ve mba hanombanana ny fahaiza-mamorona?'
+{"step": 1, "gsm8k_n": 256, "gsm8k_quick_chat": 0.59375, "t_eval_s": 16.7}
+{"step": 2, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.589608316678511, "tokens": 185370, "cumulative_loss_tokens": 360026, "grad_norm": 4.53125, "lr": 9e-06, "finish_rate": 0.984, "comp_len": 724.1, "dropped_truncated": 0, "gold_loss": 0.4409, "gold_lambda": 0.5, "rep_ratio": 2.349, "t_data_s": 0.0, "t_rollout_s": 55.0, "t_step_s": 121.6, "t_refresh_s": 0.3, "mem_gb": 9.93, "mem_gb_teacher": 13.51}
+The attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.
+{"step": 3, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6362397829683438, "tokens": 190611, "cumulative_loss_tokens": 550637, "grad_norm": 3.921875, "lr": 1.2e-05, "finish_rate": 0.957, "comp_len": 744.6, "dropped_truncated": 0, "gold_loss": 0.4903, "gold_lambda": 0.5, "rep_ratio": 2.642, "t_data_s": 0.0, "t_rollout_s": 56.9, "t_step_s": 124.6, "t_refresh_s": 0.3, "mem_gb": 10.23, "mem_gb_teacher": 13.62}
+{"step": 4, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5276193276248334, "tokens": 160132, "cumulative_loss_tokens": 710769, "grad_norm": 3.15625, "lr": 1.5e-05, "finish_rate": 1.0, "comp_len": 625.5, "dropped_truncated": 0, "gold_loss": 0.4785, "gold_lambda": 0.5, "rep_ratio": 2.597, "t_data_s": 0.0, "t_rollout_s": 46.6, "t_step_s": 109.9, "t_refresh_s": 0.3, "mem_gb": 9.98, "mem_gb_teacher": 13.53}
+{"step": 5, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5429938330516391, "tokens": 162385, "cumulative_loss_tokens": 873154, "grad_norm": 4.21875, "lr": 1.8e-05, "finish_rate": 0.992, "comp_len": 634.3, "dropped_truncated": 0, "gold_loss": 1.5549, "gold_lambda": 0.5, "rep_ratio": 2.847, "t_data_s": 0.0, "t_rollout_s": 49.4, "t_step_s": 110.2, "t_refresh_s": 0.3, "mem_gb": 10.26, "mem_gb_teacher": 13.62}
+{"step": 6, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5009330182323017, "tokens": 161557, "cumulative_loss_tokens": 1034711, "grad_norm": 2.875, "lr": 2.1e-05, "finish_rate": 0.984, "comp_len": 631.1, "dropped_truncated": 0, "gold_loss": 1.2, "gold_lambda": 0.5, "rep_ratio": 2.285, "t_data_s": 0.0, "t_rollout_s": 51.3, "t_step_s": 110.1, "t_refresh_s": 0.3, "mem_gb": 10.08, "mem_gb_teacher": 13.56}
+{"step": 7, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.4394795861535276, "tokens": 179710, "cumulative_loss_tokens": 1214421, "grad_norm": 2.40625, "lr": 2.4e-05, "finish_rate": 0.996, "comp_len": 702.0, "dropped_truncated": 0, "gold_loss": 1.56, "gold_lambda": 0.5, "rep_ratio": 2.334, "t_data_s": 0.0, "t_rollout_s": 54.2, "t_step_s": 118.6, "t_refresh_s": 0.3, "mem_gb": 9.93, "mem_gb_teacher": 13.53}
+{"step": 8, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.46311661079864236, "tokens": 172672, "cumulative_loss_tokens": 1387093, "grad_norm": 2.375, "lr": 2.7000000000000002e-05, "finish_rate": 0.992, "comp_len": 674.5, "dropped_truncated": 0, "gold_loss": 1.6159, "gold_lambda": 0.5, "rep_ratio": 2.589, "t_data_s": 0.0, "t_rollout_s": 50.6, "t_step_s": 112.0, "t_refresh_s": 0.3, "mem_gb": 10.01, "mem_gb_teacher": 13.54}
+{"step": 9, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5894883901328609, "tokens": 177365, "cumulative_loss_tokens": 1564458, "grad_norm": 1.46875, "lr": 3e-05, "finish_rate": 0.98, "comp_len": 692.8, "dropped_truncated": 0, "gold_loss": 0.49, "gold_lambda": 0.5, "rep_ratio": 2.318, "t_data_s": 0.0, "t_rollout_s": 53.0, "t_step_s": 116.8, "t_refresh_s": 0.3, "mem_gb": 10.16, "mem_gb_teacher": 13.59}
+{"step": 10, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5003410960365811, "tokens": 179180, "cumulative_loss_tokens": 1743638, "grad_norm": 1.2109375, "lr": 3e-05, "finish_rate": 0.98, "comp_len": 699.9, "dropped_truncated": 0, "gold_loss": 0.4765, "gold_lambda": 0.5, "rep_ratio": 2.69, "t_data_s": 0.0, "t_rollout_s": 53.2, "t_step_s": 116.4, "t_refresh_s": 0.3, "mem_gb": 9.91, "mem_gb_teacher": 13.5}
+[eval step 10] sample: 'Misy, Fapsana, and Mbaumona are three different words. Misy refers to a a noun, Fapsana is a noun, and Mbaumona is a noun. The word Mamba is a noun, and -on is a suffix that turns the word into a noun'
+{"step": 11, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.4788465677366256, "tokens": 169222, "cumulative_loss_tokens": 1912860, "grad_norm": 7.65625, "lr": 3e-05, "finish_rate": 0.988, "comp_len": 661.0, "dropped_truncated": 0, "gold_loss": 1.8904, "gold_lambda": 0.5, "rep_ratio": 2.661, "t_data_s": 0.0, "t_rollout_s": 50.4, "t_step_s": 112.0, "t_refresh_s": 0.3, "mem_gb": 9.92, "mem_gb_teacher": 13.51}
+{"step": 12, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.4898799755617264, "tokens": 190041, "cumulative_loss_tokens": 2102901, "grad_norm": 5.0, "lr": 3e-05, "finish_rate": 0.984, "comp_len": 742.3, "dropped_truncated": 0, "gold_loss": 1.6883, "gold_lambda": 0.5, "rep_ratio": 2.72, "t_data_s": 0.0, "t_rollout_s": 55.1, "t_step_s": 121.7, "t_refresh_s": 0.3, "mem_gb": 10.29, "mem_gb_teacher": 13.63}
+{"step": 13, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6191565852231934, "tokens": 179426, "cumulative_loss_tokens": 2282327, "grad_norm": 3.453125, "lr": 3e-05, "finish_rate": 0.973, "comp_len": 700.9, "dropped_truncated": 0, "gold_loss": 1.7587, "gold_lambda": 0.5, "rep_ratio": 2.516, "t_data_s": 0.0, "t_rollout_s": 51.9, "t_step_s": 114.8, "t_refresh_s": 0.3, "mem_gb": 9.91, "mem_gb_teacher": 13.49}
+{"step": 14, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5307344244249613, "tokens": 173105, "cumulative_loss_tokens": 2455432, "grad_norm": 4.125, "lr": 3e-05, "finish_rate": 0.973, "comp_len": 676.2, "dropped_truncated": 0, "gold_loss": 1.681, "gold_lambda": 0.5, "rep_ratio": 2.729, "t_data_s": 0.0, "t_rollout_s": 51.0, "t_step_s": 114.6, "t_refresh_s": 0.3, "mem_gb": 9.86, "mem_gb_teacher": 13.48}
+{"step": 15, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6075845100702332, "tokens": 182124, "cumulative_loss_tokens": 2637556, "grad_norm": 2.34375, "lr": 3e-05, "finish_rate": 0.992, "comp_len": 711.4, "dropped_truncated": 0, "gold_loss": 0.4579, "gold_lambda": 0.5, "rep_ratio": 2.474, "t_data_s": 0.0, "t_rollout_s": 53.4, "t_step_s": 117.2, "t_refresh_s": 0.3, "mem_gb": 9.92, "mem_gb_teacher": 13.51}
+{"step": 16, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5922139764446475, "tokens": 180837, "cumulative_loss_tokens": 2818393, "grad_norm": 2.078125, "lr": 3e-05, "finish_rate": 0.973, "comp_len": 706.4, "dropped_truncated": 0, "gold_loss": 0.4464, "gold_lambda": 0.5, "rep_ratio": 2.593, "t_data_s": 0.0, "t_rollout_s": 52.7, "t_step_s": 119.9, "t_refresh_s": 0.3, "mem_gb": 9.91, "mem_gb_teacher": 13.51}
+{"step": 17, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6971930358045317, "tokens": 168901, "cumulative_loss_tokens": 2987294, "grad_norm": 2.09375, "lr": 3e-05, "finish_rate": 0.969, "comp_len": 659.8, "dropped_truncated": 0, "gold_loss": 1.3389, "gold_lambda": 0.5, "rep_ratio": 2.484, "t_data_s": 0.0, "t_rollout_s": 51.5, "t_step_s": 116.2, "t_refresh_s": 0.3, "mem_gb": 10.15, "mem_gb_teacher": 13.59}
+{"step": 18, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6623637577042752, "tokens": 178358, "cumulative_loss_tokens": 3165652, "grad_norm": 1.875, "lr": 3e-05, "finish_rate": 0.984, "comp_len": 696.7, "dropped_truncated": 0, "gold_loss": 1.3371, "gold_lambda": 0.5, "rep_ratio": 2.515, "t_data_s": 0.0, "t_rollout_s": 55.5, "t_step_s": 122.1, "t_refresh_s": 0.3, "mem_gb": 10.43, "mem_gb_teacher": 13.68}
+{"step": 19, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6558139320164235, "tokens": 180673, "cumulative_loss_tokens": 3346325, "grad_norm": 1.671875, "lr": 3e-05, "finish_rate": 0.965, "comp_len": 705.8, "dropped_truncated": 0, "gold_loss": 1.593, "gold_lambda": 0.5, "rep_ratio": 2.497, "t_data_s": 0.0, "t_rollout_s": 54.9, "t_step_s": 121.6, "t_refresh_s": 0.3, "mem_gb": 10.12, "mem_gb_teacher": 13.57}
+{"step": 20, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5545611710471838, "tokens": 198365, "cumulative_loss_tokens": 3544690, "grad_norm": 2.609375, "lr": 3e-05, "finish_rate": 0.965, "comp_len": 774.9, "dropped_truncated": 0, "gold_loss": 1.6895, "gold_lambda": 0.5, "rep_ratio": 2.682, "t_data_s": 0.0, "t_rollout_s": 59.5, "t_step_s": 130.4, "t_refresh_s": 0.3, "mem_gb": 10.21, "mem_gb_teacher": 13.61}
+[eval step 20] sample: 'Misy mba hambonana ny fahaiz-mamanona ny se mba hambonana ny fahaiz-mamanona ny se mba hambonana ny fahaiz-mamanona ny se mba hambonana ny fahaiz-mamanona'
+{"step": 20, "gsm8k_n": 256, "gsm8k_quick_chat": 0.57421875, "t_eval_s": 21.4}
+{"step": 21, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7081326860186451, "tokens": 153659, "cumulative_loss_tokens": 3698349, "grad_norm": 2.59375, "lr": 3e-05, "finish_rate": 0.969, "comp_len": 600.2, "dropped_truncated": 0, "gold_loss": 0.3889, "gold_lambda": 0.5, "rep_ratio": 2.61, "t_data_s": 0.0, "t_rollout_s": 47.0, "t_step_s": 108.7, "t_refresh_s": 0.3, "mem_gb": 10.0, "mem_gb_teacher": 13.54}
+{"step": 22, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6475236291024463, "tokens": 175549, "cumulative_loss_tokens": 3873898, "grad_norm": 1.8671875, "lr": 3e-05, "finish_rate": 0.984, "comp_len": 685.7, "dropped_truncated": 0, "gold_loss": 0.3969, "gold_lambda": 0.5, "rep_ratio": 2.524, "t_data_s": 0.0, "t_rollout_s": 54.1, "t_step_s": 123.2, "t_refresh_s": 0.3, "mem_gb": 10.13, "mem_gb_teacher": 13.58}
+{"step": 23, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7012895279056457, "tokens": 156583, "cumulative_loss_tokens": 4030481, "grad_norm": 1.9921875, "lr": 3e-05, "finish_rate": 0.973, "comp_len": 611.7, "dropped_truncated": 0, "gold_loss": 0.9875, "gold_lambda": 0.5, "rep_ratio": 2.474, "t_data_s": 0.0, "t_rollout_s": 49.0, "t_step_s": 112.0, "t_refresh_s": 0.3, "mem_gb": 9.95, "mem_gb_teacher": 13.51}
+{"step": 24, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.8830423523515076, "tokens": 142499, "cumulative_loss_tokens": 4172980, "grad_norm": 9.5625, "lr": 3e-05, "finish_rate": 0.988, "comp_len": 556.6, "dropped_truncated": 0, "gold_loss": 1.126, "gold_lambda": 0.5, "rep_ratio": 2.431, "t_data_s": 0.0, "t_rollout_s": 45.3, "t_step_s": 103.6, "t_refresh_s": 0.3, "mem_gb": 10.08, "mem_gb_teacher": 13.57}
+{"step": 25, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6655911572447009, "tokens": 185008, "cumulative_loss_tokens": 4357988, "grad_norm": 2.921875, "lr": 3e-05, "finish_rate": 0.953, "comp_len": 722.7, "dropped_truncated": 0, "gold_loss": 1.9257, "gold_lambda": 0.5, "rep_ratio": 2.694, "t_data_s": 0.0, "t_rollout_s": 56.8, "t_step_s": 125.6, "t_refresh_s": 0.3, "mem_gb": 10.33, "mem_gb_teacher": 13.65}
+{"step": 26, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.5897857546141282, "tokens": 172974, "cumulative_loss_tokens": 4530962, "grad_norm": 2.03125, "lr": 3e-05, "finish_rate": 0.973, "comp_len": 675.7, "dropped_truncated": 0, "gold_loss": 1.6761, "gold_lambda": 0.5, "rep_ratio": 2.771, "t_data_s": 0.0, "t_rollout_s": 52.4, "t_step_s": 118.5, "t_refresh_s": 0.3, "mem_gb": 10.27, "mem_gb_teacher": 13.63}
+{"step": 27, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6517713041402011, "tokens": 172075, "cumulative_loss_tokens": 4703037, "grad_norm": 2.171875, "lr": 3e-05, "finish_rate": 0.977, "comp_len": 672.2, "dropped_truncated": 0, "gold_loss": 1.2191, "gold_lambda": 0.5, "rep_ratio": 2.356, "t_data_s": 0.0, "t_rollout_s": 51.2, "t_step_s": 115.7, "t_refresh_s": 0.3, "mem_gb": 10.03, "mem_gb_teacher": 13.54}
+{"step": 28, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.8185856890662986, "tokens": 129102, "cumulative_loss_tokens": 4832139, "grad_norm": 4.03125, "lr": 3e-05, "finish_rate": 0.988, "comp_len": 504.3, "dropped_truncated": 0, "gold_loss": 1.333, "gold_lambda": 0.5, "rep_ratio": 2.367, "t_data_s": 0.0, "t_rollout_s": 42.8, "t_step_s": 99.2, "t_refresh_s": 0.3, "mem_gb": 10.03, "mem_gb_teacher": 13.54}
+{"step": 29, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7672473138648033, "tokens": 144369, "cumulative_loss_tokens": 4976508, "grad_norm": 4.65625, "lr": 3e-05, "finish_rate": 0.961, "comp_len": 563.9, "dropped_truncated": 0, "gold_loss": 0.5047, "gold_lambda": 0.5, "rep_ratio": 2.525, "t_data_s": 0.0, "t_rollout_s": 46.2, "t_step_s": 110.5, "t_refresh_s": 0.3, "mem_gb": 10.15, "mem_gb_teacher": 13.59}
+{"step": 30, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7387997474785865, "tokens": 146453, "cumulative_loss_tokens": 5122961, "grad_norm": 3.546875, "lr": 3e-05, "finish_rate": 0.988, "comp_len": 572.1, "dropped_truncated": 0, "gold_loss": 0.4805, "gold_lambda": 0.5, "rep_ratio": 2.373, "t_data_s": 0.0, "t_rollout_s": 46.5, "t_step_s": 109.7, "t_refresh_s": 0.3, "mem_gb": 10.26, "mem_gb_teacher": 13.62}
+[eval step 30] sample: 'Misya, as a a term, signifies a condition of being in a state of being in a state of being in a state of being in a state of being in a state of being in a state of being in a state of being in a stat'
+{"step": 31, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.6948283791479429, "tokens": 184172, "cumulative_loss_tokens": 5307133, "grad_norm": 4.15625, "lr": 3e-05, "finish_rate": 0.945, "comp_len": 719.4, "dropped_truncated": 0, "gold_loss": 1.624, "gold_lambda": 0.5, "rep_ratio": 2.974, "t_data_s": 0.0, "t_rollout_s": 54.5, "t_step_s": 123.4, "t_refresh_s": 0.3, "mem_gb": 9.99, "mem_gb_teacher": 13.53}
+{"step": 32, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7520270073016101, "tokens": 161634, "cumulative_loss_tokens": 5468767, "grad_norm": 3.125, "lr": 3e-05, "finish_rate": 0.941, "comp_len": 631.4, "dropped_truncated": 0, "gold_loss": 1.385, "gold_lambda": 0.5, "rep_ratio": 2.371, "t_data_s": 0.0, "t_rollout_s": 51.0, "t_step_s": 122.5, "t_refresh_s": 0.3, "mem_gb": 10.15, "mem_gb_teacher": 13.59}
+{"step": 33, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7959626553265123, "tokens": 174953, "cumulative_loss_tokens": 5643720, "grad_norm": 6.625, "lr": 3e-05, "finish_rate": 0.941, "comp_len": 683.4, "dropped_truncated": 0, "gold_loss": 1.7494, "gold_lambda": 0.5, "rep_ratio": 2.384, "t_data_s": 0.0, "t_rollout_s": 54.0, "t_step_s": 125.1, "t_refresh_s": 0.3, "mem_gb": 10.12, "mem_gb_teacher": 13.57}
+{"step": 34, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.683204715838983, "tokens": 197798, "cumulative_loss_tokens": 5841518, "grad_norm": 12.4375, "lr": 3e-05, "finish_rate": 0.941, "comp_len": 772.6, "dropped_truncated": 0, "gold_loss": 2.2089, "gold_lambda": 0.5, "rep_ratio": 2.54, "t_data_s": 0.0, "t_rollout_s": 59.9, "t_step_s": 132.0, "t_refresh_s": 0.3, "mem_gb": 9.99, "mem_gb_teacher": 13.53}
+{"step": 35, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7359022620276259, "tokens": 224154, "cumulative_loss_tokens": 6065672, "grad_norm": 2.53125, "lr": 3e-05, "finish_rate": 0.895, "comp_len": 875.6, "dropped_truncated": 0, "gold_loss": 0.3891, "gold_lambda": 0.5, "rep_ratio": 2.576, "t_data_s": 0.0, "t_rollout_s": 66.7, "t_step_s": 146.6, "t_refresh_s": 0.3, "mem_gb": 10.4, "mem_gb_teacher": 13.67}
+{"step": 36, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.995374442305058, "tokens": 238556, "cumulative_loss_tokens": 6304228, "grad_norm": 3.0, "lr": 3e-05, "finish_rate": 0.855, "comp_len": 931.9, "dropped_truncated": 0, "gold_loss": 0.4685, "gold_lambda": 0.5, "rep_ratio": 3.499, "t_data_s": 0.0, "t_rollout_s": 75.3, "t_step_s": 159.4, "t_refresh_s": 0.3, "mem_gb": 10.2, "mem_gb_teacher": 13.6}
+{"step": 37, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.8644908860862507, "tokens": 221896, "cumulative_loss_tokens": 6526124, "grad_norm": 10.8125, "lr": 3e-05, "finish_rate": 0.871, "comp_len": 866.8, "dropped_truncated": 0, "gold_loss": 1.5978, "gold_lambda": 0.5, "rep_ratio": 2.379, "t_data_s": 0.0, "t_rollout_s": 66.9, "t_step_s": 148.7, "t_refresh_s": 0.3, "mem_gb": 9.9, "mem_gb_teacher": 13.5}
+{"step": 38, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 0.7681629302413664, "tokens": 270641, "cumulative_loss_tokens": 6796765, "grad_norm": 15.0, "lr": 3e-05, "finish_rate": 0.84, "comp_len": 1057.2, "dropped_truncated": 0, "gold_loss": 1.5569, "gold_lambda": 0.5, "rep_ratio": 2.879, "t_data_s": 0.0, "t_rollout_s": 99.2, "t_step_s": 188.3, "t_refresh_s": 0.3, "mem_gb": 10.18, "mem_gb_teacher": 13.59}
+{"step": 39, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 1.1499663096357071, "tokens": 262364, "cumulative_loss_tokens": 7059129, "grad_norm": 9.0, "lr": 3e-05, "finish_rate": 0.781, "comp_len": 1024.9, "dropped_truncated": 0, "gold_loss": 0.4787, "gold_lambda": 0.5, "rep_ratio": 4.284, "t_data_s": 0.0, "t_rollout_s": 96.1, "t_step_s": 181.6, "t_refresh_s": 0.3, "mem_gb": 10.0, "mem_gb_teacher": 13.53}
+{"step": 40, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 1.1472117115308085, "tokens": 270179, "cumulative_loss_tokens": 7329308, "grad_norm": 9.6875, "lr": 3e-05, "finish_rate": 0.777, "comp_len": 1055.4, "dropped_truncated": 0, "gold_loss": 0.484, "gold_lambda": 0.5, "rep_ratio": 3.411, "t_data_s": 0.0, "t_rollout_s": 101.5, "t_step_s": 192.6, "t_refresh_s": 0.3, "mem_gb": 10.21, "mem_gb_teacher": 13.6}
+[eval step 40] sample: "I can summarize the relationship between the two people and the progress made so far.\n\nLet's start with the basics and work our way towards the end.\n\n1. **Introduction and Basic Relationship:**\n\nThe f"
+{"step": 40, "gsm8k_n": 256, "gsm8k_quick_chat": 0.5546875, "t_eval_s": 42.1}
+{"step": 41, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 1.289868695305495, "tokens": 307658, "cumulative_loss_tokens": 7636966, "grad_norm": 22.125, "lr": 3e-05, "finish_rate": 0.707, "comp_len": 1201.8, "dropped_truncated": 0, "gold_loss": 1.7162, "gold_lambda": 0.5, "rep_ratio": 7.153, "t_data_s": 0.0, "t_rollout_s": 111.7, "t_step_s": 202.1, "t_refresh_s": 0.3, "mem_gb": 10.13, "mem_gb_teacher": 13.57}
+{"step": 42, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 1.597719869015183, "tokens": 325402, "cumulative_loss_tokens": 7962368, "grad_norm": 34.25, "lr": 3e-05, "finish_rate": 0.684, "comp_len": 1271.1, "dropped_truncated": 0, "gold_loss": 1.8621, "gold_lambda": 0.5, "rep_ratio": 3.49, "t_data_s": 0.0, "t_rollout_s": 117.8, "t_step_s": 209.5, "t_refresh_s": 0.3, "mem_gb": 10.22, "mem_gb_teacher": 13.69}
+{"step": 43, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 3.4306665449676728, "tokens": 468403, "cumulative_loss_tokens": 8430771, "grad_norm": 736.0, "lr": 3e-05, "finish_rate": 0.168, "comp_len": 1829.7, "dropped_truncated": 0, "gold_loss": 5.2979, "gold_lambda": 0.5, "rep_ratio": 1.949, "t_data_s": 0.0, "t_rollout_s": 190.4, "t_step_s": 289.5, "t_refresh_s": 0.3, "mem_gb": 10.43, "mem_gb_teacher": 13.68}
+{"step": 44, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 4.171818237882794, "tokens": 505709, "cumulative_loss_tokens": 8936480, "grad_norm": 1064.0, "lr": 3e-05, "finish_rate": 0.066, "comp_len": 1975.4, "dropped_truncated": 0, "gold_loss": 7.5318, "gold_lambda": 0.5, "rep_ratio": 1.675, "t_data_s": 0.0, "t_rollout_s": 204.4, "t_step_s": 303.7, "t_refresh_s": 0.3, "mem_gb": 10.22, "mem_gb_teacher": 13.6}
+{"step": 45, "epoch": 0, "training_mode": "on-policy", "reverse_kl": 4.106944052248165, "tokens": 489030, "cumulative_loss_tokens": 9425510, "grad_norm": 980.0, "lr": 3e-05, "finish_rate": 0.074, "comp_len": 1910.3, "dropped_truncated": 0, "gold_loss": 6.428, "gold_lambda": 0.5, "rep_ratio": 1.691, "t_data_s": 0.0, "t_rollout_s": 196.6, "t_step_s": 296.7, "t_refresh_s": 0.3, "mem_gb": 10.15, "mem_gb_teacher": 13.58}

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/files/requirements.txt

@@ -0,0 +1,130 @@
+certifi==2026.6.17
+fonttools==4.63.0
+urllib3==2.7.0
+requests==2.34.2
+nvidia-cublas-cu12==12.6.4.1
+packaging==26.2
+regex==2026.6.28
+portalocker==3.2.0
+safetensors==0.8.0
+datasets==5.0.0
+sqlitedict==2.1.0
+narwhals==2.23.0
+latex2sympy2_extended==1.11.0
+smmap==5.0.3
+colorama==0.4.6
+anyio==4.14.1
+cycler==0.12.1
+pytablewriter==1.2.1
+pydantic_core==2.46.4
+huggingface_hub==0.36.2
+charset-normalizer==3.4.8
+nvidia-cusolver-cu12==11.7.1.2
+joblib==1.5.3
+typing-inspection==0.4.2
+wandb==0.28.0
+bitsandbytes==0.49.2
+kiwisolver==1.5.0
+cuda-toolkit==12.6.3
+propcache==0.5.2
+Jinja2==3.1.6
+cuda-bindings==12.9.7
+pluggy==1.6.0
+pytest==9.1.1
+fsspec==2026.4.0
+tqdm==4.68.3
+psutil==7.2.2
+pydantic==2.13.4
+typepy==1.3.5
+xxhash==3.8.1
+DataProperty==1.1.1
+aiosignal==1.4.0
+annotated-types==0.7.0
+Pygments==2.20.0
+zstandard==0.25.0
+threadpoolctl==3.6.0
+liger_kernel==0.8.1
+nvidia-cudnn-cu12==9.10.2.21
+evaluate==0.4.6
+pillow==12.3.0
+GitPython==3.1.50
+httpx==0.28.1
+lm_eval==0.4.12
+pytz==2026.2
+mbstrdecoder==1.1.5
+immutabledict==4.3.1
+dill==0.4.1
+mpmath==1.3.0
+yarl==1.24.2
+word2number==1.1
+networkx==3.6.1
+einops==0.8.2
+tabulate==0.10.0
+pathvalidate==3.3.1
+tenacity==9.1.4
+sentry-sdk==2.64.0
+nvidia-nvshmem-cu12==3.4.5
+hf-xet==1.5.1
+six==1.17.0
+pyarrow==24.0.0
+nvidia-curand-cu12==10.3.7.77
+cuda-pathfinder==1.5.6
+sacrebleu==2.6.0
+triton==3.7.1
+frozenlist==1.8.0
+antlr4-python3-runtime==4.11.0
+tabledata==1.3.5
+h11==0.16.0
+nvidia-cusparselt-cu12==0.7.1
+nvidia-nvjitlink-cu12==12.6.85
+scipy==1.18.0
+rouge_score==0.1.2
+transformers==4.57.6
+multiprocess==0.70.19
+absl-py==2.5.0
+accelerate==1.14.0
+click==8.4.2
+lxml==6.1.1
+defusedxml==0.7.1
+setuptools==81.0.0
+nvidia-cuda-cupti-cu12==12.6.80
+httpcore==1.0.9
+langdetect==1.0.9
+nltk==3.10.0
+iniconfig==2.3.0
+nvidia-cuda-runtime-cu12==12.6.77
+nvidia-cusparse-cu12==12.5.4.2
+matplotlib==3.11.0
+protobuf==7.35.1
+nvidia-cufft-cu12==11.3.0.4
+pyparsing==3.3.2
+chardet==6.0.0.post1
+typing_extensions==4.16.0
+stanford-stk==0.7.1
+aiohappyeyeballs==2.7.1
+MarkupSafe==3.0.3
+nvidia-cufile-cu12==1.11.1.6
+tcolorpy==0.1.7
+scikit-learn==1.9.0
+tokenizers==0.22.2
+attrs==26.1.0
+gitdb==4.0.12
+aiohttp==3.14.1
+glean==0.0.1
+platformdirs==4.10.0
+megablocks==0.11.0.dev0
+nvidia-cuda-nvrtc-cu12==12.6.85
+pandas==3.0.3
+multidict==6.7.1
+torch==2.12.1+cu126
+contourpy==1.3.3
+nvidia-nvtx-cu12==12.6.77
+math-verify==0.9.0
+filelock==3.29.5
+idna==3.18
+nvidia-nccl-cu12==2.29.3
+python-dateutil==2.9.0.post0
+more-itertools==11.1.0
+numpy==2.0.2
+sympy==1.14.0
+PyYAML==6.0.3

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/files/wandb-metadata.json

@@ -0,0 +1 @@
+{"os": "Linux-7.0.11-76070011-generic-x86_64-with-glibc2.35", "python": "CPython 3.12.12", "started_at": "2026-08-01T19:14:40.480723Z", "args": ["--student", "outputs/pruned/glean-0125inst-math-keep50", "--teacher", "allenai/OLMoE-1B-7B-0125-Instruct", "--training-mode", "on-policy", "--kl-direction", "reverse", "--dataset", "allenai/Dolci-Instruct-RL", "--lr", "3e-5", "--optimizer", "adamw8bit", "--epochs", "2", "--sweep", "60", "--max-new-tokens", "2048", "--group-size", "4", "--prompts-per-step", "256", "--micro-batch", "4", "--seed", "1223", "--save-every", "1000", "--gsm8k-every", "20", "--gsm8k-max-new-tokens", "1024", "--fast-teacher", "--liger-loss", "--gold-mix-lambda", "0.5", "--gold-loss", "ce", "--gold-topk-targets", "outputs/teacher_trajectories/dolci_combined_top128", "--rollout-engine", "vllm", "--vllm-gpu", "2", "--vllm-refresh-every", "1", "--student-device", "cuda:1", "--teacher-device", "cuda:0", "--out-dir", "outputs/healed/mixceonly_keep50", "--wandb", "--wandb-project", "glean-heal", "--wandb-run-name", "mixce-keep50-s1223"], "program": "/home/henry/Documents/PythonProjects/variable-reap/scripts/11_distill_on_policy.py", "code_path": "scripts/11_distill_on_policy.py", "code_path_local": "scripts/11_distill_on_policy.py", "git": {"remote_url": "https://github.com/hbfreed/variable-reap.git", "commit": "22365a35ebf321b7ccfa3430a715a153e787d3ee"}, "root": "outputs/healed/mixceonly_keep50", "host": "pop-os", "executable": "/home/henry/Documents/PythonProjects/variable-reap/.venv/bin/python", "cpu_count": 24, "cpu_count_logical": 48, "gpu_type": "NVIDIA GeForce RTX 3090", "gpu_count": 3, "disk": {"/": {"total": "1958315118592", "used": "1293356351488"}}, "memory": {"total": "134898364416"}, "gpu_nvidia": [{"name": "NVIDIA GeForce RTX 3090", "memory_total": "25769803776", "cuda_cores": 10496, "architecture": "Ampere", "uuid": "GPU-8ca70870-ddf2-d274-bcc5-182c2075bced"}, {"name": "NVIDIA GeForce RTX 3090", "memory_total": "25769803776", "cuda_cores": 10496, "architecture": "Ampere", "uuid": "GPU-a6acf07f-31f5-618f-a5d0-c0017e7e2e27"}, {"name": "NVIDIA GeForce RTX 3090", "memory_total": "25769803776", "cuda_cores": 10496, "architecture": "Ampere", "uuid": "GPU-864c54df-0130-7780-e271-8a5551d1733f"}], "cuda_version": "13.0", "writer_id": "cdl6ysspj1b4fap16js7r9iidg0igxex"}

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/files/wandb-summary.json

@@ -0,0 +1 @@
+{"_runtime": 6464.45969866, "reverse_kl": 4.106944052248165, "rep_ratio": 1.691, "mem_gb_teacher": 13.58, "_step": 47, "cumulative_loss_tokens": 9425510, "grad_norm": 980, "lr": 3e-05, "gold_lambda": 0.5, "t_data_s": 0, "t_rollout_s": 196.6, "_timestamp": 1785618145.529177, "step": 45, "epoch": 0, "training_mode": "on-policy", "tokens": 489030, "comp_len": 1910.3, "t_refresh_s": 0.3, "finish_rate": 0.074, "dropped_truncated": 0, "gold_loss": 6.428, "t_step_s": 296.7, "mem_gb": 10.15, "gsm8k_quick_chat": 0.5546875}

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug-core.log

@@ -0,0 +1,11 @@
+{"time":"2026-08-01T12:14:40.743385653-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpfogj7ky0/port-13469.txt","pid":13469,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-08-01T12:14:40.744231298-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":13469}
+{"time":"2026-08-01T12:14:40.744203577-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-13469-13592-2475494762/socket","Net":"unix"}}
+{"time":"2026-08-01T12:14:40.921493362-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-08-01T12:14:40.930576146-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"umzrxo81","id":"1(@)"}
+{"time":"2026-08-01T12:14:41.051163293-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"umzrxo81","id":"1(@)"}
+{"time":"2026-08-01T14:10:07.26454867-07:00","level":"INFO","msg":"connection: closing","id":"1(@)"}
+{"time":"2026-08-01T14:10:07.264684404-07:00","level":"INFO","msg":"connection: closed successfully","id":"1(@)"}
+{"time":"2026-08-01T14:10:07.264561551-07:00","level":"INFO","msg":"processOutgoingData: finished","id":"1(@)"}
+{"time":"2026-08-01T14:10:07.264738306-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"1(@)"}
+{"time":"2026-08-01T14:10:07.565864468-07:00","level":"INFO","msg":"server: parent process exited, terminating service process"}

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug-internal.log

@@ -0,0 +1,11 @@
+{"time":"2026-08-01T12:14:40.930763762-07:00","level":"INFO","msg":"wandb-core"}
+{"time":"2026-08-01T12:14:40.930810353-07:00","level":"INFO","msg":"stream: starting","core version":"0.28.0"}
+{"time":"2026-08-01T12:14:41.050876465-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T12:14:41.050913746-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T12:14:41.050952917-07:00","level":"INFO","msg":"stream: created new stream","id":"umzrxo81"}
+{"time":"2026-08-01T12:14:41.05106738-07:00","level":"INFO","msg":"handler: started"}
+{"time":"2026-08-01T12:14:41.051150532-07:00","level":"INFO","msg":"stream: started"}
+{"time":"2026-08-01T12:14:41.051193074-07:00","level":"INFO","msg":"writer: started","stream_id":"umzrxo81"}
+{"time":"2026-08-01T12:14:41.051238735-07:00","level":"INFO","msg":"sender: started"}
+{"time":"2026-08-01T12:14:41.065739707-07:00","level":"WARN","msg":"featurechecker: GraphQL client is nil, skipping feature loading"}
+{"time":"2026-08-01T12:14:41.065767307-07:00","level":"WARN","msg":"runupserter: server does not expand metric globs but the x_server_side_expand_glob_metrics setting is set; ignoring"}

healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug.log

@@ -0,0 +1,19 @@
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_setup.py:_flush():81] Configure stats pid to 13469
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:setup_run_log_directory():725] Logging user logs to outputs/healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug.log
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to outputs/healed/mixceonly_keep50/wandb/offline-run-20260801_121440-umzrxo81/logs/debug-internal.log
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:init():768] calling init triggers
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+config: {'student': 'outputs/pruned/glean-0125inst-math-keep50', 'teacher': 'allenai/OLMoE-1B-7B-0125-Instruct', 'training_mode': 'on-policy', 'kl_direction': 'reverse', 'dataset': 'allenai/Dolci-Instruct-RL', 'dataset_sources': None, 'max_difficulty': None, 'trajectories': 'outputs/teacher_trajectories/dolci_math_curated.jsonl', 'trajectory_dataset': 'allenai/Dolci-Instruct-RL', 'off_policy_frames': 'chat', 'off_policy_max_seq_len': 2048, 'topk_targets': None, 'max_loss_tokens': None, 'loss_tokens_per_step': None, 'teacher_device': 'cuda:0', 'student_device': 'cuda:1', 'lr': 3e-05, 'optimizer': 'adamw8bit', 'weight_decay': 0.1, 'epochs': 2, 'prompts_per_step': 256, 'group_size': 4, 'rollout_batch': 64, 'micro_batch': 4, 'max_new_tokens': 2048, 'max_prompt_len': 1024, 'warmup_steps': 10, 'max_grad_norm': 1.0, 'eval_every': 10, 'gsm8k_every': 20, 'gsm8k_n': 256, 'gsm8k_batch': 16, 'gsm8k_max_new_tokens': 1024, 'gsm8k_frames': 'chat', 'save_every': 1000, 'out_dir': 'outputs/healed/mixceonly_keep50', 'sweep': 60, 'wandb': True, 'wandb_project': 'glean-heal', 'wandb_run_name': 'mixce-keep50-s1223', 'wandb_run_id': None, 'wandb_resume': None, 'wandb_mode': 'offline', 'no_wandb_sync': False, 'debug': False, 'resume_from': None, 'start_step': 0, 'no_grad_checkpointing': False, 'seed': 1223, 'no_teacher_overlap': False, 'sync_checkpoints': False, 'rollout_engine': 'vllm', 'vllm_gpu': '2', 'vllm_port': 8377, 'vllm_refresh_every': 1, 'vllm_serve_bin': 'vllm-plugin/.venv/bin/python', 'vllm_gpu_mem_util': 0.85, 'liger_loss': True, 'gold_mix_lambda': 0.5, 'gold_topk_targets': 'outputs/teacher_trajectories/dolci_combined_top128', 'gold_loss': 'ce', 'gold_mix_decay': 0.0, 'fast_teacher': True, 'reference_kl_beta': 0.0, 'drop_truncated_rollouts': False, 'vllm_max_model_len': None, 'vllm_refresh_mode': 'reload', 'vllm_live_dir': None, 'resolved_kl_direction': 'reverse', '_wandb': {}}
+2026-08-01 12:14:40,482 INFO    MainThread:13469 [wandb_init.py:init():816] starting backend
+2026-08-01 12:14:40,921 INFO    MainThread:13469 [wandb_init.py:init():831] sending inform_init request
+2026-08-01 12:14:41,052 INFO    MainThread:13469 [wandb_init.py:init():836] backend started and connected
+2026-08-01 12:14:41,054 INFO    MainThread:13469 [wandb_init.py:init():906] updated telemetry
+2026-08-01 12:14:41,062 INFO    MainThread:13469 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
+2026-08-01 12:14:41,068 INFO    MainThread:13469 [wandb_init.py:init():974] starting run threads in backend
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_console_start():2523] atexit reg
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_redirect():2373] redirect: wrap_raw
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_redirect():2442] Wrapping output streams.
+2026-08-01 12:14:41,197 INFO    MainThread:13469 [wandb_run.py:_redirect():2465] Redirects installed.
+2026-08-01 12:14:41,199 INFO    MainThread:13469 [wandb_init.py:init():1012] run started, returning control to user process

healed/opd_warm_keep50/vllm_live/chat_template.jinja

@@ -0,0 +1,9 @@
+{{ bos_token }}{% for message in messages %}{% if message['role'] == 'system' %}{{ '<|system|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'user' %}{{ '<|user|>
+' + message['content'] + '
+' }}{% elif message['role'] == 'assistant' %}{% if not loop.last %}{{ '<|assistant|>
+'  + message['content'] + eos_token + '
+' }}{% else %}{{ '<|assistant|>
+'  + message['content'] + eos_token }}{% endif %}{% endif %}{% if loop.last and add_generation_prompt %}{{ '<|assistant|>
+' }}{% endif %}{% endfor %}

healed/opd_warm_keep50/vllm_live/config.json

@@ -0,0 +1,887 @@
+{
+  "architectures": [
+    "PrunedOlmoeForCausalLM"
+  ],
+  "attention_bias": false,
+  "attention_dropout": 0.0,
+  "auto_map": {
+    "AutoConfig": "configuration_pruned_olmoe.PrunedOlmoeConfig",
+    "AutoModelForCausalLM": "modeling_pruned_olmoe.PrunedOlmoeForCausalLM"
+  },
+  "clip_qkv": null,
+  "dtype": "bfloat16",
+  "eos_token_id": 50279,
+  "expert_widths": [
+    [
+      1024,
+      384,
+      256,
+      256,
+      768,
+      1024,
+      128,
+      768,
+      384,
+      128,
+      1024,
+      384,
+      384,
+      128,
+      640,
+      896,
+      896,
+      256,
+      768,
+      128,
+      768,
+      768,
+      768,
+      768,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      256,
+      768,
+      512,
+      640,
+      768,
+      896,
+      1024,
+      384,
+      384,
+      512,
+      384,
+      256,
+      768,
+      640,
+      896,
+      896,
+      512,
+      768,
+      512,
+      384,
+      1024,
+      896,
+      896,
+      768,
+      128
+    ],
+    [
+      768,
+      896,
+      896,
+      256,
+      896,
+      384,
+      768,
+      768,
+      1024,
+      384,
+      512,
+      1024,
+      768,
+      768,
+      1024,
+      640,
+      128,
+      768,
+      384,
+      768,
+      128,
+      256,
+      512,
+      384,
+      896,
+      256,
+      1024,
+      1024,
+      896,
+      1024,
+      896,
+      768,
+      256,
+      768,
+      128,
+      1024,
+      768,
+      768,
+      640,
+      384,
+      1024,
+      384,
+      640,
+      896,
+      1024,
+      512,
+      512,
+      512,
+      640
+    ],
+    [
+      768,
+      1024,
+      512,
+      384,
+      384,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      640,
+      256,
+      1024,
+      1024,
+      1024,
+      640,
+      1024,
+      896,
+      512,
+      768,
+      128,
+      768,
+      1024,
+      768,
+      256,
+      1024,
+      896,
+      1024,
+      896,
+      512,
+      640,
+      384,
+      896,
+      384,
+      512,
+      640,
+      256,
+      768,
+      256,
+      768,
+      768,
+      640,
+      1024,
+      640,
+      896,
+      768,
+      768,
+      256
+    ],
+    [
+      640,
+      640,
+      896,
+      512,
+      640,
+      896,
+      896,
+      384,
+      896,
+      896,
+      384,
+      896,
+      640,
+      256,
+      640,
+      1024,
+      1024,
+      768,
+      1024,
+      896,
+      1024,
+      768,
+      896,
+      256,
+      512,
+      768,
+      1024,
+      256,
+      768,
+      512,
+      256,
+      640,
+      1024,
+      1024,
+      512,
+      1024,
+      768,
+      256,
+      1024,
+      384,
+      896,
+      1024,
+      896,
+      1024,
+      384
+    ],
+    [
+      256,
+      640,
+      640,
+      896,
+      768,
+      896,
+      768,
+      768,
+      896,
+      896,
+      1024,
+      256,
+      512,
+      1024,
+      640,
+      896,
+      512,
+      512,
+      384,
+      384,
+      256,
+      384,
+      384,
+      896,
+      896,
+      768,
+      640,
+      896,
+      768,
+      1024,
+      512,
+      640,
+      512,
+      640,
+      896,
+      512,
+      512,
+      384,
+      640,
+      896,
+      896,
+      896,
+      1024,
+      640,
+      1024,
+      640,
+      1024
+    ],
+    [
+      1024,
+      512,
+      1024,
+      1024,
+      640,
+      896,
+      640,
+      1024,
+      896,
+      384,
+      1024,
+      128,
+      896,
+      768,
+      1024,
+      768,
+      640,
+      896,
+      768,
+      640,
+      512,
+      896,
+      512,
+      640,
+      256,
+      768,
+      640,
+      768,
+      384,
+      896,
+      512,
+      512,
+      256,
+      512,
+      896,
+      256,
+      384,
+      640,
+      512,
+      640,
+      896,
+      512,
+      1024,
+      256,
+      768,
+      1024,
+      768,
+      256,
+      256
+    ],
+    [
+      640,
+      896,
+      1024,
+      896,
+      1024,
+      1024,
+      1024,
+      512,
+      256,
+      256,
+      1024,
+      768,
+      512,
+      768,
+      1024,
+      1024,
+      1024,
+      384,
+      512,
+      1024,
+      512,
+      1024,
+      128,
+      640,
+      640,
+      896,
+      768,
+      128,
+      256,
+      256,
+      256,
+      256,
+      896,
+      1024,
+      1024,
+      384,
+      896,
+      256,
+      896,
+      640,
+      1024,
+      384,
+      640,
+      256,
+      1024,
+      1024,
+      1024
+    ],
+    [
+      1024,
+      384,
+      1024,
+      1024,
+      256,
+      128,
+      256,
+      384,
+      256,
+      384,
+      896,
+      768,
+      896,
+      896,
+      512,
+      896,
+      640,
+      384,
+      384,
+      896,
+      768,
+      384,
+      896,
+      768,
+      768,
+      512,
+      896,
+      768,
+      768,
+      896,
+      128,
+      896,
+      512,
+      256,
+      768,
+      128,
+      384,
+      256,
+      896,
+      896,
+      384,
+      768,
+      512,
+      640,
+      256,
+      768,
+      640,
+      896,
+      384,
+      512,
+      1024,
+      768,
+      384
+    ],
+    [
+      512,
+      768,
+      512,
+      256,
+      128,
+      640,
+      384,
+      640,
+      768,
+      896,
+      640,
+      768,
+      256,
+      384,
+      1024,
+      896,
+      256,
+      896,
+      512,
+      256,
+      896,
+      768,
+      256,
+      896,
+      896,
+      384,
+      896,
+      640,
+      768,
+      512,
+      768,
+      768,
+      1024,
+      768,
+      640,
+      768,
+      384,
+      256,
+      512,
+      896,
+      128,
+      384,
+      256,
+      768,
+      384,
+      256,
+      1024,
+      1024,
+      896,
+      256,
+      1024,
+      256,
+      128,
+      896
+    ],
+    [
+      640,
+      640,
+      896,
+      256,
+      1024,
+      512,
+      1024,
+      768,
+      384,
+      512,
+      256,
+      768,
+      896,
+      768,
+      512,
+      768,
+      768,
+      640,
+      384,
+      768,
+      512,
+      768,
+      768,
+      512,
+      768,
+      128,
+      896,
+      512,
+      768,
+      1024,
+      128,
+      384,
+      768,
+      768,
+      768,
+      384,
+      512,
+      640,
+      768,
+      512,
+      768,
+      1024,
+      640,
+      896,
+      256,
+      1024,
+      384,
+      768,
+      768,
+      768
+    ],
+    [
+      896,
+      512,
+      896,
+      768,
+      384,
+      384,
+      768,
+      512,
+      768,
+      512,
+      1024,
+      640,
+      896,
+      896,
+      256,
+      640,
+      1024,
+      256,
+      896,
+      128,
+      128,
+      128,
+      768,
+      896,
+      384,
+      896,
+      512,
+      896,
+      384,
+      256,
+      640,
+      640,
+      896,
+      768,
+      640,
+      256,
+      896,
+      896,
+      512,
+      128,
+      896,
+      256,
+      256,
+      640,
+      896,
+      896,
+      128,
+      1024,
+      256,
+      384,
+      1024,
+      640,
+      896
+    ],
+    [
+      768,
+      384,
+      640,
+      896,
+      256,
+      128,
+      384,
+      896,
+      128,
+      128,
+      896,
+      256,
+      384,
+      896,
+      512,
+      768,
+      768,
+      512,
+      512,
+      768,
+      896,
+      640,
+      768,
+      896,
+      896,
+      896,
+      640,
+      896,
+      640,
+      512,
+      896,
+      256,
+      512,
+      128,
+      512,
+      384,
+      768,
+      768,
+      1024,
+      256,
+      768,
+      256,
+      768,
+      512,
+      640,
+      1024,
+      128,
+      896,
+      896,
+      896,
+      896,
+      1024
+    ],
+    [
+      512,
+      384,
+      768,
+      640,
+      640,
+      768,
+      1024,
+      896,
+      512,
+      256,
+      640,
+      768,
+      640,
+      896,
+      128,
+      256,
+      896,
+      1024,
+      256,
+      640,
+      512,
+      256,
+      128,
+      512,
+      256,
+      640,
+      768,
+      768,
+      128,
+      128,
+      768,
+      640,
+      1024,
+      1024,
+      768,
+      512,
+      896,
+      768,
+      896,
+      1024,
+      896,
+      896,
+      512,
+      640,
+      1024,
+      512,
+      1024,
+      512,
+      512,
+      512,
+      768
+    ],
+    [
+      896,
+      896,
+      1024,
+      1024,
+      896,
+      128,
+      768,
+      256,
+      1024,
+      256,
+      1024,
+      640,
+      384,
+      256,
+      256,
+      512,
+      768,
+      896,
+      512,
+      768,
+      384,
+      1024,
+      896,
+      896,
+      1024,
+      896,
+      768,
+      896,
+      640,
+      1024,
+      512,
+      896,
+      512,
+      1024,
+      512,
+      512,
+      256,
+      256,
+      256,
+      512,
+      768,
+      128,
+      384,
+      512,
+      896,
+      896,
+      1024,
+      256,
+      384,
+      640
+    ],
+    [
+      896,
+      640,
+      384,
+      512,
+      256,
+      640,
+      1024,
+      384,
+      1024,
+      1024,
+      768,
+      256,
+      1024,
+      768,
+      512,
+      896,
+      256,
+      1024,
+      768,
+      768,
+      768,
+      384,
+      384,
+      256,
+      1024,
+      384,
+      384,
+      384,
+      896,
+      768,
+      640,
+      768,
+      512,
+      896,
+      896,
+      896,
+      896,
+      256,
+      384,
+      128,
+      1024,
+      896,
+      256,
+      256,
+      768,
+      640,
+      896,
+      384,
+      768,
+      512,
+      640
+    ],
+    [
+      896,
+      1024,
+      768,
+      1024,
+      896,
+      256,
+      768,
+      128,
+      128,
+      768,
+      512,
+      896,
+      384,
+      768,
+      1024,
+      256,
+      768,
+      768,
+      256,
+      512,
+      512,
+      640,
+      512,
+      256,
+      768,
+      896,
+      384,
+      1024,
+      640,
+      1024,
+      512,
+      512,
+      384,
+      512,
+      512,
+      1024,
+      384,
+      896,
+      768,
+      384,
+      384,
+      128,
+      384,
+      1024,
+      896,
+      640,
+      768,
+      768,
+      256,
+      640,
+      512,
+      640,
+      384
+    ]
+  ],
+  "glean_metadata": {
+    "base_model": "allenai/OLMoE-1B-7B-0125-Instruct",
+    "block_size": 128,
+    "criterion": "reap",
+    "dead_experts": 217,
+    "keep_fraction": 0.5,
+    "min_width": 128,
+    "params": 3697491968,
+    "scores": "outputs/scores_0125inst_dolmino-math/scores.pt"
+  },
+  "hidden_act": "silu",
+  "hidden_size": 2048,
+  "initializer_range": 0.02,
+  "intermediate_size": 1024,
+  "max_position_embeddings": 4096,
+  "model_type": "pruned_olmoe",
+  "norm_topk_prob": false,
+  "num_attention_heads": 16,
+  "num_experts": 64,
+  "num_experts_per_tok": 8,
+  "num_hidden_layers": 16,
+  "num_key_value_heads": 16,
+  "output_router_logits": false,
+  "pad_token_id": 1,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 10000.0,
+  "router_aux_loss_coef": 0.01,
+  "tie_word_embeddings": false,
+  "transformers_version": "4.57.6",
+  "use_cache": false,
+  "vocab_size": 50304
+}

healed/opd_warm_keep50/vllm_live/configuration_pruned_olmoe.py

@@ -0,0 +1,27 @@
+"""Configuration for GLEAN-pruned OLMoE: variable-width, variable-count experts.
+"""
+
+from transformers.models.olmoe.configuration_olmoe import OlmoeConfig
+
+
+class PrunedOlmoeConfig(OlmoeConfig):
+    """OlmoeConfig plus a per-(layer, expert) width table.
+
+    ``expert_widths[l]`` lists the SwiGLU intermediate width of each surviving
+    expert in decoder layer ``l``, in expert order. Lists are ragged: layers
+    may keep different numbers of experts (deleted experts simply don't
+    appear — the router in layer ``l`` has ``len(expert_widths[l])`` rows),
+    and each width may differ (multiples of the GEMM block size, 128, for
+    variable-MegaBlocks execution). ``None`` means an unpruned model
+    (uniform ``num_experts`` × ``intermediate_size``).
+
+    The inherited ``num_experts`` / ``intermediate_size`` keep their ORIGINAL
+    (pre-pruning) values for provenance; the width table is authoritative for
+    the built architecture.
+    """
+
+    model_type = "pruned_olmoe"
+
+    def __init__(self, expert_widths: list[list[int]] | None = None, **kwargs):
+        super().__init__(**kwargs)
+        self.expert_widths = expert_widths