"""Kernel-level speedups for p-image-oss (the diffusers fork) on an RTX PRO 6000 Blackwell (sm_120). No torch.compile. * `install_adaln_tables`: every block's AdaLN projection (13B of the transformer's 33B parameters) depends only on the timestep, and the pipeline only ever uses the timesteps of its fixed sigma grid. Their outputs are computed once per transformer call signature and the projection weights are dropped. Bitwise identical to the projections. * `fuse_qkv`: one GEMM for q, k and v. * `quantize_fp8`: W8A8 linears, per-output-channel weight scales and per-row dynamic activation scales, through torch._scaled_mm (cuBLASLt FP8). Activations are quantized by one Triton kernel per GEMM input. """ from __future__ import annotations import inspect import time import torch import triton import triton.language as tl from torch import nn FP8 = torch.float8_e4m3fn FP8_MAX = 448.0 def _free_cache() -> None: try: if torch.cuda.is_available() and torch.cuda.is_initialized(): torch.cuda.empty_cache() except Exception: # noqa: BLE001 - ZeroGPU emulates CUDA outside GPU calls pass # AdaLN tables ------------------------------------------------------------------------------------------------------- class AdaLNTable(nn.Module): """Replaces `MiniMaxH3AdaLayerNormModulation`: returns its precomputed output for the current call's timesteps.""" def __init__(self, tables: list[torch.Tensor], hidden_size: int, lookup: dict): super().__init__() for index, table in enumerate(tables): self.register_buffer(f"table_{index}", table, persistent=False) self.hidden_size = hidden_size self.lookup = lookup def forward(self, temb: torch.Tensor) -> tuple[torch.Tensor, ...]: table = getattr(self, f"table_{self.lookup['index']}") return table.view(-1, 6 * self.hidden_size).chunk(6, dim=-1) @torch.no_grad() def install_adaln_tables(transformer, calls: list[torch.Tensor]) -> int: """Precompute every block's AdaLN output for each distinct `timestep` argument in `calls` (as the pipeline passes them, one tensor per transformer call) and drop the projection weights. Each table is computed with the call's own shapes, exactly as the forward would, so the outputs are bitwise identical.""" device = next(transformer.parameters()).device keys: dict[tuple, int] = {} tembs = [] for timestep in calls: key = tuple(timestep.tolist()) if key not in keys: keys[key] = len(keys) temb = transformer.time_proj(timestep.to(device)) tembs.append(transformer.time_embedder(temb.to(transformer.time_embedder.linear_1.weight.dtype))) lookup = {"keys": keys, "index": None} freed = 0 for block in transformer.transformer_blocks: proj = block.adaln_proj dtype = proj.linear.weight.dtype tables = [proj.linear(nn.functional.silu(temb).to(dtype)).contiguous() for temb in tembs] freed += sum(p.numel() for p in proj.parameters()) block.adaln_proj = AdaLNTable(tables, proj.hidden_size, lookup) del proj def set_index(module, args, kwargs): key = tuple(kwargs["timestep"].tolist()) if key not in keys: raise ValueError(f"No AdaLN table for timesteps {key}; known: {list(keys)}") lookup["index"] = keys[key] transformer.register_forward_pre_hook(set_index, with_kwargs=True) _free_cache() return freed # QKV fusion --------------------------------------------------------------------------------------------------------- @torch.no_grad() def fuse_qkv(transformer) -> None: for block in transformer.transformer_blocks: attn = block.attn weight = torch.cat([attn.to_q.weight, attn.to_k.weight, attn.to_v.weight]) attn.to_qkv = nn.Linear(weight.shape[1], weight.shape[0], bias=False, device=weight.device, dtype=weight.dtype) attn.to_qkv.weight.copy_(weight) del attn.to_q, attn.to_k, attn.to_v attn.fused_projections = True _free_cache() # FP8 W8A8 ----------------------------------------------------------------------------------------------------------- @triton.jit def _rowwise_fp8_kernel(x_ptr, q_ptr, s_ptr, K, stride_x, stride_q, BLOCK: tl.constexpr): row = tl.program_id(0) offs = tl.arange(0, BLOCK) mask = offs < K x = tl.load(x_ptr + row * stride_x + offs, mask=mask, other=0.0).to(tl.float32) amax = tl.maximum(tl.max(tl.abs(x), axis=0), 1e-12) tl.store(q_ptr + row * stride_q + offs, (x * (448.0 / amax)).to(tl.float8e4nv), mask=mask) tl.store(s_ptr + row, amax / 448.0) def rowwise_fp8(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: """(M, K) bf16 -> (M, K) e4m3 and (M, 1) float32 scales with x ~= q * scale.""" M, K = x.shape q = torch.empty((M, K), device=x.device, dtype=FP8) scale = torch.empty((M, 1), device=x.device, dtype=torch.float32) block = triton.next_power_of_2(K) _rowwise_fp8_kernel[(M,)](x, q, scale, K, x.stride(0), q.stride(0), BLOCK=block, num_warps=8 if block <= 8192 else 16) return q, scale _W8A16_CONFIGS = [triton.Config({"BLOCK_N": n, "BLOCK_K": k}, num_warps=w, num_stages=4) for n in (16, 32, 64) for k in (128, 256) for w in (4, 8)] # Tuned configs persist in Triton's cache dir, so a new process (ZeroGPU workers) skips the autotune. _AUTOTUNE_CACHE = {"cache_results": True} if "cache_results" in inspect.signature(triton.autotune).parameters else {} @triton.autotune(configs=_W8A16_CONFIGS, key=["N", "K"], **_AUTOTUNE_CACHE) @triton.jit def _w8a16_kernel(x_ptr, w_ptr, s_ptr, b_ptr, y_ptr, M, N, K, stride_x, stride_w, stride_y, HAS_BIAS: tl.constexpr, BLOCK_N: tl.constexpr, BLOCK_K: tl.constexpr): offs_n = tl.program_id(0) * BLOCK_N + tl.arange(0, BLOCK_N) offs_m = tl.arange(0, 16) acc = tl.zeros((16, BLOCK_N), dtype=tl.float32) for k0 in range(0, K, BLOCK_K): offs_k = k0 + tl.arange(0, BLOCK_K) w = tl.load(w_ptr + offs_n[:, None] * stride_w + offs_k[None, :], mask=(offs_n[:, None] < N) & (offs_k[None, :] < K), other=0.0).to(tl.bfloat16) x = tl.load(x_ptr + offs_m[:, None] * stride_x + offs_k[None, :], mask=(offs_m[:, None] < M) & (offs_k[None, :] < K), other=0.0) acc += tl.dot(x, tl.trans(w)) acc *= tl.load(s_ptr + offs_n, mask=offs_n < N, other=0.0)[None, :] if HAS_BIAS: acc += tl.load(b_ptr + offs_n, mask=offs_n < N, other=0.0).to(tl.float32)[None, :] tl.store(y_ptr + offs_m[:, None] * stride_y + offs_n[None, :], acc.to(y_ptr.dtype.element_ty), mask=(offs_m[:, None] < M) & (offs_n[None, :] < N)) def w8a16_gemv(x: torch.Tensor, weight: torch.Tensor, scale: torch.Tensor, bias: torch.Tensor | None) -> torch.Tensor: """y = x @ (weight * scale).T for at most 16 rows of x: FP8 weights dequantized in registers, bandwidth-bound.""" M, K = x.shape N = weight.shape[0] x = x.contiguous() y = torch.empty((M, N), device=x.device, dtype=x.dtype) grid = lambda meta: (triton.cdiv(N, meta["BLOCK_N"]),) _w8a16_kernel[grid](x, weight, scale, bias if bias is not None else scale, y, M, N, K, x.stride(0), weight.stride(0), y.stride(0), HAS_BIAS=bias is not None) return y class FP8Linear(nn.Module): def __init__(self, linear: nn.Linear, fast_accum: bool = False): super().__init__() weight = linear.weight.detach() scale = weight.float().abs().amax(dim=1, keepdim=True).clamp_min(1e-12) / FP8_MAX self.register_buffer("weight", (weight.float() / scale).to(FP8).contiguous()) # (N, K) self.register_buffer("weight_scale", scale.t().contiguous()) # (1, N) self.bias = linear.bias self.in_features, self.out_features = linear.in_features, linear.out_features self.out_dtype = weight.dtype self.fast_accum = fast_accum def forward(self, x: torch.Tensor) -> torch.Tensor: shape = x.shape rows = x.reshape(-1, shape[-1]) if rows.shape[0] <= 16: # decode: weight-only, the GEMM is a bandwidth-bound GEMV return w8a16_gemv(rows.to(self.out_dtype), self.weight, self.weight_scale.view(-1), self.bias).view(*shape[:-1], self.out_features) q, scale = rowwise_fp8(rows) out = torch._scaled_mm(q, self.weight.t(), scale_a=scale, scale_b=self.weight_scale, bias=self.bias, out_dtype=self.out_dtype, use_fast_accum=self.fast_accum) return out.view(*shape[:-1], self.out_features) @torch.no_grad() def quantize_fp8(module: nn.Module, fast_accum: bool = False, skip: tuple[str, ...] = ()) -> int: """Replace every nn.Linear under `module` (names containing an entry of `skip` excluded) by an FP8Linear.""" count = 0 for name, child in list(module.named_modules()): for child_name, leaf in list(child.named_children()): full = f"{name}.{child_name}" if name else child_name if isinstance(leaf, nn.Linear) and not any(s in full for s in skip): setattr(child, child_name, FP8Linear(leaf, fast_accum)) del leaf count += 1 _free_cache() return count # CUDA-graph greedy decoding ----------------------------------------------------------------------------------------- class GraphedGreedyGenerate: """Greedy decoding of one sequence for a Qwen3_5ForConditionalGeneration, with the decode step replayed as a CUDA graph. A drop-in for `model.generate(**inputs, max_new_tokens=..., do_sample=False, eos_token_id=...)`. The prefill runs eagerly into a persistent StaticCache (its linear-attention states and KV tensors are updated in place, so their addresses never change). One decode step (embed, 32 layers, lm_head, argmax) is captured once over static buffers: the input token, its absolute position, the M-RoPE offset and the attention mask over the cache. Per token only the graph is replayed; the end token is checked every `check_every` tokens. """ def __init__(self, model, max_cache_len: int = 12288, max_new_tokens: int = 4096, check_every: int = 16): from transformers import StaticCache self.model, self.generate_eager = model, model.generate self.max_cache_len, self.check_every = max_cache_len, check_every device = model.device self.cache = StaticCache(config=model.config, max_cache_len=max_cache_len) self.tok = torch.zeros((1, 1), dtype=torch.long, device=device) self.pos = torch.zeros((1,), dtype=torch.long, device=device) self.offset = torch.zeros((4, 1, 1), dtype=torch.long, device=device) self.mask = torch.zeros((1, 1, 1, max_cache_len), dtype=torch.bool, device=device) self.out = torch.zeros((max_new_tokens,), dtype=torch.long, device=device) self.count = torch.zeros((1,), dtype=torch.long, device=device) self.graph = None def _step(self): text_model = self.model.model.language_model self.mask.index_fill_(3, self.pos, True) hidden = text_model( inputs_embeds=text_model.embed_tokens(self.tok), position_ids=self.pos.view(1, 1, 1) + self.offset, attention_mask={"full_attention": self.mask, "linear_attention": None}, past_key_values=self.cache, use_cache=True, ).last_hidden_state token = self.model.lm_head(hidden[:, -1]).argmax(-1) self.out.index_copy_(0, self.count, token) self.tok.copy_(token.view(1, 1)) self.pos.add_(1) self.count.add_(1) def _prefill(self, input_ids, kwargs): self.cache.reset() self.mask.zero_() self.model.model.rope_deltas = None logits = self.model(input_ids=input_ids, past_key_values=self.cache, use_cache=True, logits_to_keep=1, **kwargs).logits token = logits[:, -1].argmax(-1) rope_deltas = self.model.model.rope_deltas self.offset.zero_() if rope_deltas is not None: self.offset[1:] = rope_deltas.view(-1)[0] prompt_length = input_ids.shape[1] self.mask[..., :prompt_length] = True self.out[0] = token[0] self.count.fill_(1) self.tok.copy_(token.view(1, 1)) self.pos.fill_(prompt_length) def _capture(self): stream = torch.cuda.Stream() stream.wait_stream(torch.cuda.current_stream()) with torch.cuda.stream(stream): for _ in range(3): self._step() torch.cuda.current_stream().wait_stream(stream) self.graph = torch.cuda.CUDAGraph() with torch.cuda.graph(self.graph): self._step() @torch.no_grad() def generate(self, input_ids, max_new_tokens: int, eos_token_id: int, **kwargs): kwargs.pop("do_sample", None) kwargs.pop("pad_token_id", None) if input_ids.shape[0] != 1 or input_ids.shape[1] + max_new_tokens > self.max_cache_len: return self.generate_eager(input_ids=input_ids, max_new_tokens=max_new_tokens, do_sample=False, eos_token_id=eos_token_id, **kwargs) if self.graph is None: self._prefill(input_ids, kwargs) self._capture() self._prefill(input_ids, kwargs) generated = 1 while generated < max_new_tokens and not bool((self.out[:generated] == eos_token_id).any()): for _ in range(min(self.check_every, max_new_tokens - generated)): self.graph.replay() generated = min(generated + self.check_every, max_new_tokens) tokens = self.out[:generated] ends = (tokens == eos_token_id).nonzero() if ends.numel(): tokens = tokens[: int(ends[0, 0]) + 1] return torch.cat([input_ids, tokens[None].to(input_ids.device)], dim=1) # MTP speculative decoding ------------------------------------------------------------------------------------------- class MTPHead(nn.Module): """Qwen3.5's multi-token-prediction head: (hidden state at i, embedding of token i+1) -> hidden for token i+2.""" def __init__(self, config): super().__init__() from transformers.models.qwen3_5.modeling_qwen3_5 import Qwen3_5DecoderLayer, Qwen3_5RMSNorm size, eps = config.hidden_size, config.rms_norm_eps self.config = config self.fc = nn.Linear(2 * size, size, bias=False) self.pre_fc_norm_embedding = Qwen3_5RMSNorm(size, eps) self.pre_fc_norm_hidden = Qwen3_5RMSNorm(size, eps) self.layers = nn.ModuleList([Qwen3_5DecoderLayer(config, 0)]) self.norm = Qwen3_5RMSNorm(size, eps) def forward(self, embeds, hidden, position_embeddings, text_positions, mask, cache): x = self.fc(torch.cat([self.pre_fc_norm_embedding(embeds), self.pre_fc_norm_hidden(hidden)], dim=-1)) x = self.layers[0](x, position_embeddings=position_embeddings, attention_mask=mask, position_ids=text_positions, past_key_values=cache, use_cache=True) return self.norm(x) MTP_REPO, MTP_REVISION = "Qwen/Qwen3.5-9B", "c202236235762e1c871ad0ccb60c8ee5ba337b9a" def read_hub_tensors(repo: str, revision: str, prefix: str) -> dict[str, torch.Tensor]: """The tensors named `prefix*` of a sharded safetensors repo, read by byte range (no full shard downloads).""" import json import struct from huggingface_hub import HfFileSystem, hf_hub_download dtypes = {"BF16": torch.bfloat16, "F16": torch.float16, "F32": torch.float32} index = json.load(open(hf_hub_download(repo, "model.safetensors.index.json", revision=revision))) fs, tensors = HfFileSystem(), {} for file in sorted({f for name, f in index["weight_map"].items() if name.startswith(prefix)}): with fs.open(f"{repo}/{file}", "rb", revision=revision) as handle: header_size = struct.unpack(" MTPHead: """Qwen3.5-9B's MTP head (transformers drops it), used as the prompt upsampler's draft model.""" import copy config = copy.deepcopy(model.config.get_text_config()) config.num_hidden_layers, config.layer_types = 1, ["full_attention"] state = {name[len("mtp."):]: tensor for name, tensor in read_hub_tensors(repo, revision, "mtp.").items()} head = MTPHead(config) head.load_state_dict(state) return head.to(model.device, model.dtype).eval() GDN_VERIFY = {"torch_reference": False} def _gdn_verify(module, hidden_states, cache_params): """Qwen3_5GatedDeltaNet over a few cached tokens, one token at a time, keeping the conv and recurrent states after every token in `module._spec_snaps` so a rejected draft rolls back by index.""" from torch.nn import functional as F from transformers.models.qwen3_5 import modeling_qwen3_5 as qm batch, length, _ = hidden_states.shape layer = cache_params.layers[module.layer_idx] conv_state, recurrent_state = layer.conv_states[0], layer.recurrent_states[0] conv_snaps, recurrent_snaps = module._spec_snaps mixed = module.in_proj_qkv(hidden_states).transpose(1, 2) z = module.in_proj_z(hidden_states).reshape(batch, length, -1, module.head_v_dim) b, a = module.in_proj_b(hidden_states), module.in_proj_a(hidden_states) convolved = [] for j in range(length): convolved.append(qm.causal_conv1d_update(mixed[:, :, j : j + 1], conv_state, module.conv1d.weight.squeeze(1), module.conv1d.bias, module.activation)) conv_snaps[j].copy_(conv_state) mixed = torch.cat(convolved, dim=2).transpose(1, 2) query, key, value = torch.split(mixed, [module.key_dim, module.key_dim, module.value_dim], dim=-1) query = query.reshape(batch, length, -1, module.head_k_dim) key = key.reshape(batch, length, -1, module.head_k_dim) value = value.reshape(batch, length, -1, module.head_v_dim) beta = b.sigmoid() g = -module.A_log.float().exp() * F.softplus(a.float() + module.dt_bias) if module.num_v_heads // module.num_k_heads > 1: query = query.repeat_interleave(module.num_v_heads // module.num_k_heads, dim=2) key = key.repeat_interleave(module.num_v_heads // module.num_k_heads, dim=2) recurrence = qm.torch_recurrent_gated_delta_rule if GDN_VERIFY["torch_reference"]: recurrence = getattr(recurrence, "__wrapped__", recurrence) outputs = [] for j in range(length): step = [t[:, j : j + 1].contiguous() for t in (query, key, value, g, beta)] out, state = recurrence(step[0], step[1], step[2], g=step[3], beta=step[4], initial_state=recurrent_state.clone(), output_final_state=True, use_qk_l2norm_in_kernel=True) recurrent_state.copy_(state) recurrent_snaps[j].copy_(recurrent_state) outputs.append(out) core = module.norm(torch.cat(outputs, dim=1).reshape(-1, module.head_v_dim), z.reshape(-1, module.head_v_dim)) return module.out_proj(core.reshape(batch, length, -1)) def _install_gdn_verify(module) -> None: """Route `module`'s forward through `_gdn_verify` while `module._spec_verify` is set. Patched on the instance: `set_use_kernels` installs instance-level forwards, which a class-level patch would not reach.""" if getattr(module, "_spec_patched", False): return original = module.forward def forward(hidden_states, cache_params=None, attention_mask=None, **kwargs): if module._spec_verify: return _gdn_verify(module, hidden_states, cache_params) return original(hidden_states, cache_params=cache_params, attention_mask=attention_mask, **kwargs) module.forward, module._spec_verify, module._spec_patched = forward, False, True class MTPGreedyGenerate: """Greedy decoding of one sequence with one MTP draft token per step, each step one CUDA graph. Lossless: every emitted token is the main model's own greedy choice; an accepted draft adds a second token to the step. A step feeds `[t, draft]` to the main model (its Gated DeltaNet layers keep their states after each token, the full-attention layers write both K/V rows), takes the greedy predictions p0, p1, accepts the draft if it equals p0, rolls the states back to the accepted prefix and emits `draft, p1` or `p0`. The MTP head then runs on the two (hidden, next token) pairs of the step and drafts from the accepted one. The end token is checked every `check_every` steps. """ def __init__(self, model, mtp: MTPHead, max_cache_len: int = 12288, max_new_tokens: int = 4096, check_every: int = 8): from transformers import StaticCache self.model, self.mtp, self.generate_eager = model, mtp, model.generate self.text_model = model.model.language_model self.max_cache_len, self.check_every = max_cache_len, check_every device = model.device self.cache = StaticCache(config=model.config, max_cache_len=max_cache_len) self.mtp_cache = StaticCache(config=mtp.config, max_cache_len=max_cache_len) self.gdn = [layer.linear_attn for layer in self.text_model.layers if layer.block_type == "linear_attention"] for module in self.gdn: _install_gdn_verify(module) long = dict(dtype=torch.long, device=device) self.tok, self.draft, self.pos = (torch.zeros((1,), **long) for _ in range(3)) self.offset = torch.zeros((4, 1, 1), **long) self.keys = torch.arange(max_cache_len, **long) self.pair = torch.arange(2, **long) self.out = torch.zeros((max_new_tokens + 2 * check_every + 2,), **long) self.count, self.accepted = torch.zeros((1,), **long), torch.zeros((1,), **long) self.max_new_tokens = max_new_tokens self.graph, self.stats = None, {} def _set_verify(self, on: bool): for module in self.gdn: module._spec_verify = on def _static_lengths(self): from transformers.cache_utils import StaticLayer return [layer.cumulative_length for cache in (self.cache, self.mtp_cache) for layer in cache.layers if isinstance(layer, StaticLayer)] def _step(self): positions = self.pos + self.pair position_ids = positions.view(1, 1, 2) + self.offset mask = (self.keys.view(1, -1) <= positions.view(-1, 1)).view(1, 1, 2, -1) masks = {"full_attention": mask, "linear_attention": None} tokens = torch.cat([self.tok, self.draft]).view(1, 2) hidden = self.text_model(inputs_embeds=self.text_model.embed_tokens(tokens), position_ids=position_ids, attention_mask=masks, past_key_values=self.cache, use_cache=True).last_hidden_state preds = self.model.lm_head(hidden).argmax(-1).view(2) accept = (preds[0] == self.draft[0]).long().view(1) for module in self.gdn: layer = self.cache.layers[module.layer_idx] conv_snaps, recurrent_snaps = module._spec_snaps layer.conv_states[0].copy_(conv_snaps.index_select(0, accept)[0]) layer.recurrent_states[0].copy_(recurrent_snaps.index_select(0, accept)[0]) emitted = torch.stack([torch.where(accept.bool(), self.draft, preds[:1])[0], preds[1]]) self.out.index_copy_(0, self.count + self.pair, emitted) cos_sin = self.text_model.rotary_emb(hidden, position_ids[1:]) mtp_hidden = self.mtp(self.text_model.embed_tokens(emitted.view(1, 2)), hidden, cos_sin, position_ids[0], mask, self.mtp_cache) draft = self.model.lm_head(mtp_hidden.index_select(1, accept)).argmax(-1).view(1) new_length = (self.pos + 1 + accept).view(()) for length in self._static_lengths(): length.copy_(new_length) self.tok.copy_(torch.where(accept.bool(), preds[1:], preds[:1])) self.draft.copy_(draft) self.pos.add_(1 + accept) self.count.add_(1 + accept) self.accepted.add_(accept) def _prefill(self, input_ids, kwargs): self.cache.reset() self.mtp_cache.reset() self._set_verify(False) self.model.model.rope_deltas = None seen = {} hook = self.text_model.register_forward_pre_hook( lambda module, args, kw: seen.update(position_ids=kw.get("position_ids")), with_kwargs=True) hidden = self.model.model(input_ids=input_ids, past_key_values=self.cache, use_cache=True, **kwargs).last_hidden_state hook.remove() first = self.model.lm_head(hidden[:, -1]).argmax(-1) length = input_ids.shape[1] steps = torch.arange(length, device=input_ids.device) rope = seen.get("position_ids") rope = steps.view(1, 1, -1).expand(3, 1, -1) if rope is None else rope[-3:] mask = (self.keys.view(1, -1) <= steps.view(-1, 1)).view(1, 1, length, -1) next_ids = torch.cat([input_ids[:, 1:], first.view(1, 1)], dim=1) mtp_hidden = self.mtp(self.text_model.embed_tokens(next_ids), hidden, self.text_model.rotary_emb(hidden, rope), steps.view(1, -1), mask, self.mtp_cache) draft = self.model.lm_head(mtp_hidden[:, -1]).argmax(-1) rope_deltas = self.model.model.rope_deltas self.offset.zero_() if rope_deltas is not None: self.offset[1:] = rope_deltas.view(-1)[0] self.out[0] = first[0] self.count.fill_(1) self.accepted.zero_() self.tok.copy_(first) self.draft.copy_(draft) self.pos.fill_(length) self._set_verify(True) def _capture(self): for module in self.gdn: layer = self.cache.layers[module.layer_idx] module._spec_snaps = (torch.empty((2, *layer.conv_states[0].shape), dtype=layer.conv_states[0].dtype, device=layer.conv_states[0].device), torch.empty((2, *layer.recurrent_states[0].shape), dtype=layer.recurrent_states[0].dtype, device=layer.recurrent_states[0].device)) stream = torch.cuda.Stream() stream.wait_stream(torch.cuda.current_stream()) with torch.cuda.stream(stream): for _ in range(3): self._step() torch.cuda.current_stream().wait_stream(stream) self.graph = torch.cuda.CUDAGraph() with torch.cuda.graph(self.graph): self._step() @torch.no_grad() def generate(self, input_ids, max_new_tokens: int, eos_token_id: int, **kwargs): kwargs.pop("do_sample", None) kwargs.pop("pad_token_id", None) margin = 2 * self.check_every + 2 self._set_verify(False) if (input_ids.shape[0] != 1 or max_new_tokens > self.max_new_tokens or input_ids.shape[1] + max_new_tokens + margin > self.max_cache_len): return self.generate_eager(input_ids=input_ids, max_new_tokens=max_new_tokens, do_sample=False, eos_token_id=eos_token_id, **kwargs) if self.graph is None: self._prefill(input_ids, kwargs) self._capture() self._prefill(input_ids, kwargs) steps, count = 0, 1 while count < max_new_tokens and not bool((self.out[:count] == eos_token_id).any()): for _ in range(self.check_every): self.graph.replay() steps += self.check_every count = int(self.count) tokens = self.out[: min(count, max_new_tokens)] ends = (tokens == eos_token_id).nonzero() if ends.numel(): tokens = tokens[: int(ends[0, 0]) + 1] self._set_verify(False) self.stats = {"steps": steps, "tokens": int(tokens.numel()), "accepted": int(self.accepted), "acceptance": round(int(self.accepted) / max(steps, 1), 3)} return torch.cat([input_ids, tokens[None].to(input_ids.device)], dim=1) # Decode attention --------------------------------------------------------------------------------------------------- def install_decode_attention(*configs) -> None: """Full attention for a few query rows over a long static cache as two grouped batched GEMVs (K/V read once per KV head, no repeat). SDPA's memory-efficient kernel parallelizes a single query poorly. Longer queries (prefill) keep SDPA.""" from transformers import AttentionInterface from transformers.integrations.sdpa_attention import sdpa_attention_forward def attention(module, query, key, value, attention_mask, dropout=0.0, scaling=None, **kwargs): batch, heads, length, dim = query.shape if length > 16 or attention_mask is None or attention_mask.dtype != torch.bool: return sdpa_attention_forward(module, query, key, value, attention_mask, dropout=dropout, scaling=scaling, **kwargs) kv_heads, keys = key.shape[1], key.shape[2] groups = heads // kv_heads q = query.reshape(batch, kv_heads, groups * length, dim) scores = torch.matmul(q, key.transpose(-1, -2)).float().view(batch, kv_heads, groups, length, keys) scores = scores * (scaling if scaling is not None else dim**-0.5) scores = scores.masked_fill(~attention_mask[..., :keys].view(1, 1, 1, length, keys), float("-inf")) probs = scores.softmax(-1).to(value.dtype).view(batch, kv_heads, groups * length, keys) out = torch.matmul(probs, value).view(batch, heads, length, -1) return out.transpose(1, 2).contiguous(), None AttentionInterface.register("p_image_oss_decode", attention) for config in configs: config._attn_implementation = "p_image_oss_decode" # Batched VAE tile decode -------------------------------------------------------------------------------------------- def batch_vae_tiles(vae, max_tiles: int = 96) -> None: """Decode the spatial tiles of `AutoencoderKLPH3Image` in batches of up to `max_tiles` instead of one decoder call per tile. The tile layout, the per-tile decoder math and the stitching are unchanged.""" def decode_clip(z: torch.Tensor) -> torch.Tensor: if not vae.use_tiling: return vae.decoder(vae.post_quant_conv(z)) ratio = vae.spatial_compression_ratio y_indices, y_lengths, y_overlaps = vae._split_tiles( z.shape[-2] * ratio, vae.tile_sample_min_height, vae.tile_sample_min_overlap_height) x_indices, x_lengths, x_overlaps = vae._split_tiles( z.shape[-1] * ratio, vae.tile_sample_min_width, vae.tile_sample_min_overlap_width) tiles = {} for i, (i_pos, i_len) in enumerate(zip(y_indices, y_lengths)): for j, (j_pos, j_len) in enumerate(zip(x_indices, x_lengths)): tiles[(i, j)] = z[..., i_pos // ratio: i_pos // ratio + i_len // ratio, j_pos // ratio: j_pos // ratio + j_len // ratio] decoded = {} by_shape: dict[tuple, list] = {} for key, tile in tiles.items(): by_shape.setdefault(tuple(tile.shape), []).append(key) for keys in by_shape.values(): for start in range(0, len(keys), max_tiles): chunk = keys[start: start + max_tiles] out = vae.decoder(vae.post_quant_conv(torch.cat([tiles[k] for k in chunk], dim=0))) for index, key in enumerate(chunk): decoded[key] = out[index: index + 1] rows = [[decoded[(i, j)] for j in range(len(x_indices))] for i in range(len(y_indices))] return vae._stitch_tiles(rows, y_overlaps, x_overlaps) vae._decode_clip = decode_clip # The Space's pipeline ------------------------------------------------------------------------------------------------ def load_slim_pipeline(path: str, adaln_calls: list, device: str = "cuda", log=print): """p-image-oss from a local snapshot, slimmed for a 48 GB slice. Components are loaded and slimmed one at a time (bf16 transformer peak ~66 GB host RAM) and moved to `device`. Returns the pipeline and the upsampler's MTP head.""" import gc from diffusers import ModularPipeline start = time.time() pipe = ModularPipeline.from_pretrained(path) pipe.load_components(names=["transformer"], dtype=torch.bfloat16) install_adaln_tables(pipe.transformer, [torch.tensor(call) for call in adaln_calls]) quantize_fp8(pipe.transformer.transformer_blocks) fuse_dit_blocks(pipe.transformer) pipe.transformer.to(device) gc.collect() log(f"[{time.time() - start:6.1f}s] transformer ready") pipe.load_components(names=["prompt_upsampler", "prompt_upsampler_processor"], dtype=torch.bfloat16) upsampler = pipe.prompt_upsampler quantize_fp8(upsampler.model.language_model) upsampler.lm_head = FP8Linear(upsampler.lm_head) mtp = load_mtp_head(upsampler) quantize_fp8(mtp) install_decode_attention(upsampler.model.language_model.config, mtp.config) upsampler.to(device) mtp.to(device) gc.collect() log(f"[{time.time() - start:6.1f}s] prompt upsampler ready") pipe.load_components(dtype=torch.bfloat16) quantize_fp8(pipe.text_encoder.model.language_model) quantize_fp8(pipe.vae.decoder.transformer_blocks) batch_vae_tiles(pipe.vae) pipe.to(device) gc.collect() log(f"[{time.time() - start:6.1f}s] pipeline ready") return pipe, mtp # Fused DiT block ---------------------------------------------------------------------------------------------------- # # One MiniMaxH3TransformerBlock in five row kernels around the FP8 GEMMs and attention. Every producer of a GEMM input # holds the whole row, so it also computes the row's FP8 scale and writes e4m3 directly (no separate quant pass): # norm1 + AdaLN modulation -> FP8 | QKV GEMM | q/k RMSNorm + RoPE | attention | out GEMM # gated residual + norm2 + modulation -> FP8 | FF-in GEMM | SwiGLU -> FP8 | FF-out GEMM | gated residual # Math runs in fp32 and rounds once per output (eager rounds to bf16 after every op). @triton.jit def _rms_mod_fp8_kernel(x_ptr, w_ptr, scale_ptr, shift_ptr, idx_ptr, q_ptr, s_ptr, n_cols, stride_x, stride_tab, eps, BLOCK: tl.constexpr): row = tl.program_id(0).to(tl.int64) cols = tl.arange(0, BLOCK) mask = cols < n_cols x = tl.load(x_ptr + row * stride_x + cols, mask=mask, other=0.0).to(tl.float32) w = tl.load(w_ptr + cols, mask=mask, other=0.0).to(tl.float32) tab = tl.load(idx_ptr + row).to(tl.int64) * stride_tab + cols sc = tl.load(scale_ptr + tab, mask=mask, other=0.0).to(tl.float32) sh = tl.load(shift_ptr + tab, mask=mask, other=0.0).to(tl.float32) y = x * tl.math.rsqrt(tl.sum(x * x, axis=0) / n_cols + eps) * w * (1.0 + sc) + sh amax = tl.maximum(tl.max(tl.where(mask, tl.abs(y), 0.0), axis=0), 1e-12) tl.store(q_ptr + row * n_cols + cols, (y * (448.0 / amax)).to(tl.float8e4nv), mask=mask) tl.store(s_ptr + row, amax / 448.0) @triton.jit def _res_gate_rms_mod_fp8_kernel(h_ptr, b_ptr, gate_ptr, w_ptr, scale_ptr, shift_ptr, idx_ptr, out_ptr, q_ptr, s_ptr, n_cols, stride_h, stride_b, stride_tab, eps, BLOCK: tl.constexpr): row = tl.program_id(0).to(tl.int64) cols = tl.arange(0, BLOCK) mask = cols < n_cols tab = tl.load(idx_ptr + row).to(tl.int64) * stride_tab + cols h = tl.load(h_ptr + row * stride_h + cols, mask=mask, other=0.0).to(tl.float32) b = tl.load(b_ptr + row * stride_b + cols, mask=mask, other=0.0).to(tl.float32) g = tl.load(gate_ptr + tab, mask=mask, other=0.0).to(tl.float32) h = (h + g * b).to(out_ptr.dtype.element_ty) tl.store(out_ptr + row * n_cols + cols, h, mask=mask) h = h.to(tl.float32) w = tl.load(w_ptr + cols, mask=mask, other=0.0).to(tl.float32) sc = tl.load(scale_ptr + tab, mask=mask, other=0.0).to(tl.float32) sh = tl.load(shift_ptr + tab, mask=mask, other=0.0).to(tl.float32) y = h * tl.math.rsqrt(tl.sum(h * h, axis=0) / n_cols + eps) * w * (1.0 + sc) + sh amax = tl.maximum(tl.max(tl.where(mask, tl.abs(y), 0.0), axis=0), 1e-12) tl.store(q_ptr + row * n_cols + cols, (y * (448.0 / amax)).to(tl.float8e4nv), mask=mask) tl.store(s_ptr + row, amax / 448.0) @triton.jit def _res_gate_kernel(h_ptr, b_ptr, gate_ptr, idx_ptr, out_ptr, n_cols, stride_h, stride_b, stride_tab, BLOCK: tl.constexpr): row = tl.program_id(0).to(tl.int64) cols = tl.arange(0, BLOCK) mask = cols < n_cols tab = tl.load(idx_ptr + row).to(tl.int64) * stride_tab + cols h = tl.load(h_ptr + row * stride_h + cols, mask=mask, other=0.0).to(tl.float32) b = tl.load(b_ptr + row * stride_b + cols, mask=mask, other=0.0).to(tl.float32) g = tl.load(gate_ptr + tab, mask=mask, other=0.0).to(tl.float32) tl.store(out_ptr + row * n_cols + cols, (h + g * b).to(out_ptr.dtype.element_ty), mask=mask) @triton.jit def _swiglu_fp8_kernel(x_ptr, q_ptr, s_ptr, n_cols, stride_x, BLOCK: tl.constexpr): row = tl.program_id(0).to(tl.int64) cols = tl.arange(0, BLOCK) mask = cols < n_cols v = tl.load(x_ptr + row * stride_x + cols, mask=mask, other=0.0).to(tl.float32) g = tl.load(x_ptr + row * stride_x + n_cols + cols, mask=mask, other=0.0).to(tl.float32) y = v * g * tl.sigmoid(g) amax = tl.maximum(tl.max(tl.where(mask, tl.abs(y), 0.0), axis=0), 1e-12) tl.store(q_ptr + row * n_cols + cols, (y * (448.0 / amax)).to(tl.float8e4nv), mask=mask) tl.store(s_ptr + row, amax / 448.0) @triton.jit def _qk_norm_rope_kernel(qkv_ptr, wq_ptr, wk_ptr, cos_ptr, sin_ptr, q_ptr, k_ptr, stride_qkv, stride_rope, heads, eps, D: tl.constexpr, ROT: tl.constexpr): row = tl.program_id(0).to(tl.int64) which = tl.program_id(1) # [0, heads) query heads, [heads, 2 * heads) key heads is_k = which >= heads head = which - heads * is_k.to(tl.int32) c = tl.arange(0, D) half = ROT // 2 partner = tl.where(c < half, c + half, tl.where(c < ROT, c - half, c)) base = qkv_ptr + row * stride_qkv + (is_k.to(tl.int64) * heads + head) * D x = tl.load(base + c).to(tl.float32) xp = tl.load(base + partner).to(tl.float32) w = tl.where(is_k, tl.load(wk_ptr + c), tl.load(wq_ptr + c)).to(tl.float32) wp = tl.where(is_k, tl.load(wk_ptr + partner), tl.load(wq_ptr + partner)).to(tl.float32) r = tl.math.rsqrt(tl.sum(x * x, axis=0) / D + eps) xn, xpn = x * r * w, xp * r * wp rot = c < ROT cos = tl.load(cos_ptr + row * stride_rope + c, mask=rot, other=1.0) sin = tl.load(sin_ptr + row * stride_rope + c, mask=rot, other=0.0) y = (xn * cos + tl.where(c < half, -xpn, xpn) * sin).to(q_ptr.dtype.element_ty) offset = (row * heads + head) * D + c tl.store(q_ptr + offset, y, mask=(c < D) & (which < heads)) tl.store(k_ptr + offset, y, mask=(c < D) & (which >= heads)) def _row_launch(kernel, rows, cols, *args, **kwargs): block = triton.next_power_of_2(cols) kernel[(rows,)](*args, BLOCK=block, num_warps=16 if block >= 16384 else 8, **kwargs) def _fp8_gemm(q, s, linear: FP8Linear): return torch._scaled_mm(q, linear.weight.t(), scale_a=s, scale_b=linear.weight_scale, bias=linear.bias, out_dtype=linear.out_dtype, use_fast_accum=linear.fast_accum) def _fused_block_forward(block, hidden_states, temb, adaln_indices, rotary_emb, attention_mask=None): from diffusers.models.attention_dispatch import dispatch_attention_fn shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = block.adaln_proj(temb) batch, rows, cols = hidden_states.shape h = hidden_states.reshape(rows, cols) idx, stride_tab, device = adaln_indices, shift_msa.stride(0), h.device attn = block.attn heads, dim = attn.heads, attn.head_dim q8 = torch.empty((rows, cols), device=device, dtype=FP8) s = torch.empty((rows, 1), device=device, dtype=torch.float32) _row_launch(_rms_mod_fp8_kernel, rows, cols, h, block.norm1.weight, scale_msa, shift_msa, idx, q8, s, cols, h.stride(0), stride_tab, block.norm1.eps) qkv = _fp8_gemm(q8, s, block._fused_qkv) q = torch.empty((rows, heads, dim), device=device, dtype=h.dtype) k = torch.empty_like(q) cos, sin = rotary_emb _qk_norm_rope_kernel[(rows, 2 * heads)](qkv, attn.norm_q.weight, attn.norm_k.weight, cos, sin, q, k, qkv.stride(0), cos.stride(0), heads, attn.norm_q.eps, D=dim, ROT=cos.shape[-1], num_warps=2) v = qkv[:, 2 * heads * dim:].view(rows, heads, dim) o = dispatch_attention_fn(q[None], k[None], v[None], attn_mask=None, dropout_p=0.0, is_causal=False, backend=attn.processor._attention_backend) attn_out = attn.to_out[0](o.reshape(rows, heads * dim)) h2 = torch.empty_like(h) _row_launch(_res_gate_rms_mod_fp8_kernel, rows, cols, h, attn_out, gate_msa, block.norm2.weight, scale_mlp, shift_mlp, idx, h2, q8, s, cols, h.stride(0), attn_out.stride(0), stride_tab, block.norm2.eps) proj = _fp8_gemm(q8, s, block.ff.net[0].proj) inner = proj.shape[1] // 2 f8 = torch.empty((rows, inner), device=device, dtype=FP8) _row_launch(_swiglu_fp8_kernel, rows, inner, proj, f8, s, inner, proj.stride(0)) ff_out = _fp8_gemm(f8, s, block.ff.net[2]) out = torch.empty_like(h) _row_launch(_res_gate_kernel, rows, cols, h2, ff_out, gate_mlp, idx, out, cols, h2.stride(0), ff_out.stride(0), stride_tab) return out.view(batch, rows, cols) @torch.no_grad() def fuse_dit_blocks(transformer) -> None: """Fused forwards for every transformer block. Needs `install_adaln_tables` and `quantize_fp8` first (and q, k, v as separate FP8Linear, which are concatenated into one QKV GEMM here).""" import types for block in transformer.transformer_blocks: attn = block.attn parts = [attn.to_q, attn.to_k, attn.to_v] fused = FP8Linear.__new__(FP8Linear) nn.Module.__init__(fused) fused.register_buffer("weight", torch.cat([p.weight for p in parts])) fused.register_buffer("weight_scale", torch.cat([p.weight_scale for p in parts], dim=1)) fused.bias, fused.out_dtype, fused.fast_accum = None, parts[0].out_dtype, parts[0].fast_accum fused.in_features, fused.out_features = parts[0].in_features, sum(p.out_features for p in parts) block._fused_qkv = fused del attn.to_q, attn.to_k, attn.to_v block.forward = types.MethodType(_fused_block_forward, block) _free_cache()