[diffusion] Fuse DiT FFN tanh-GELU into up-proj GEMM (cublasLt epilogue) behind quality=high (Qwen-Image 1024^2 denoise 12.36 -> 12.05 s on H200) (#33536)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

python/sglang/kernels/ops/diffusion/__init__.py

@@ -42,6 +42,18 @@ register_kernel(
         description="Fused residual gate-add (sglang.kernels.jit).",
     )
 )
+register_kernel(
+    KernelSpec(
+        op="diffusion.fused_linear_gelu_tanh",
+        backend=KernelBackend.TORCH,
+        target="sglang.kernels.ops.diffusion.fused_linear_gelu:fused_linear_gelu_tanh",
+        capabilities=_CUDA,
+        format_signature=FormatSignature(
+            description="linear + tanh-GELU via the cublasLt GELU epilogue"
+        ),
+        description="Fused up-proj GEMM + tanh-GELU (torch._addmm_activation).",
+    )
+)
 register_kernel(
     KernelSpec(
         op="diffusion.fused_inplace_qknorm_rope",

python/sglang/kernels/ops/diffusion/fused_linear_gelu.py

@@ -0,0 +1,182 @@
+"""Fused linear + tanh-GELU via the cublasLt GELU epilogue, gated by quality.
+
+Many diffusion DiT FeedForwards compute ``gelu(linear(x), approximate="tanh")``
+as a standalone up-projection GEMM followed by a separate, bandwidth-bound GELU
+kernel over the ``[tokens, 4*dim]`` MLP intermediate. ``torch._addmm_activation``
+folds the bias-add and GELU into the GEMM epilogue (cublasLt), removing the
+extra kernel launch and the intermediate HBM round-trip. cublasLt's GELU is the
+tanh-approximate GELU (max abs diff ~5e-6 vs ``F.gelu(approximate="tanh")`` in
+fp32), so for half-precision inference the fused path differs from the
+reference only at bf16/fp16 rounding-order level -- close, but not bit-exact.
+
+Because it is not bit-exact, the fused path is **mounted only for
+``quality="high"`` requests** (see ``SamplingParams.quality``): model code marks
+its GELU up-projection sites with :func:`mark_fused_gelu_site` (default: off,
+reference path, bit-exact), and the denoising stage calls
+:func:`mount_fused_linear_gelu` / :func:`unmount_fused_linear_gelu` at batch
+boundaries. Mounting is all-or-nothing per transformer: if any marked site
+fails the static guards (quantized weights, missing bias, non-half dtype, ...)
+no site on that transformer is fused.
+
+The fused GEMM is exposed as a registered custom op (``register_custom_op``)
+exactly like the other diffusion kernels (e.g. qknorm_rope), so it stays a
+single opaque op under ``torch.compile`` -- no graph break.
+"""
+
+from __future__ import annotations
+
+import logging
+from typing import Any, Iterator
+
+import torch
+import torch.nn as nn
+
+from sglang.srt.utils.custom_op import register_custom_op
+
+logger = logging.getLogger(__name__)
+
+# ``torch._addmm_activation`` is the (private but stable) entry point to the
+# cublasLt GEMM+bias+activation epilogue. Guard for builds where it is absent
+# so the reference path is always available.
+_HAS_ADDMM_ACTIVATION = hasattr(torch, "_addmm_activation")
+
+# Attributes of the site protocol (set by ``mark_fused_gelu_site``).
+_SITE_LINEAR_ATTR = "_sgl_fused_gelu_linear_attr"
+_SITE_ENABLED_ATTR = "_sgl_fused_gelu_enabled"
+
+
+def _fused_linear_gelu_tanh_fake(
+    x: torch.Tensor, weight: torch.Tensor, bias: torch.Tensor
+) -> torch.Tensor:
+    return x.new_empty((*x.shape[:-1], weight.shape[0]))
+
+
+@register_custom_op(
+    op_name="diffusion_fused_linear_gelu_tanh",
+    mutates_args=[],
+    fake_impl=_fused_linear_gelu_tanh_fake,
+)
+def fused_linear_gelu_tanh(
+    x: torch.Tensor, weight: torch.Tensor, bias: torch.Tensor
+) -> torch.Tensor:
+    """``gelu_tanh(x @ weight.T + bias)`` fused in the cublasLt GELU epilogue.
+
+    ``weight`` is ``[out, in]`` (nn.Linear / sglang linear layout). Registered
+    as a custom op so it is opaque under torch.compile.
+    """
+    x2d = x.reshape(-1, x.shape[-1])
+    out = torch._addmm_activation(bias, x2d, weight.t(), use_gelu=True)
+    return out.view(*x.shape[:-1], weight.shape[0])
+
+
+def _is_unquantized(linear: Any) -> bool:
+    """True iff ``linear`` carries plain, unquantized weights."""
+    # Quantized checkpoints can leave selected layers unquantized via their
+    # exclude list; keep every layer of a quantized model on the reference
+    # path (all-or-nothing would reject the model anyway).
+    if getattr(linear, "quant_config", None) is not None:
+        return False
+    quant_method = getattr(linear, "quant_method", None)
+    if quant_method is None:
+        # Plain nn.Linear has no quant_method.
+        return True
+    from sglang.multimodal_gen.runtime.layers.linear import UnquantizedLinearMethod
+
+    return isinstance(quant_method, UnquantizedLinearMethod)
+
+
+def _static_reject_reason(linear: Any) -> str | None:
+    """Why ``linear`` may never use the epilogue, or None if it may.
+
+    Input-independent guards: requires the fused API and an unquantized,
+    bias'd, non-bias-deferring linear with a half-precision 2D weight. A
+    column-parallel layer that gathers across multiple ranks is excluded (the
+    per-shard fused op would skip the cross-rank gather);
+    ``gather_output=False`` sharded layers are fine because the local shard is
+    exactly what the reference forward multiplies. The weight's *device* is
+    deliberately not checked here -- under CPU offload the weights live on CPU
+    between requests -- the runtime guard checks the input device per call.
+    """
+    if not _HAS_ADDMM_ACTIVATION:
+        return "torch._addmm_activation unavailable"
+    if not _is_unquantized(linear):
+        return "quantized linear"
+    if getattr(linear, "skip_bias_add", False):
+        return "skip_bias_add (bias returned separately)"
+    if getattr(linear, "gather_output", False) and getattr(linear, "tp_size", 1) > 1:
+        return "multi-rank gather_output"
+    weight = getattr(linear, "weight", None)
+    bias = getattr(linear, "bias", None)
+    if weight is None or bias is None or weight.dim() != 2:
+        return "missing bias or non-2D weight"
+    if weight.dtype not in (torch.bfloat16, torch.float16):
+        return f"non-half weight dtype {weight.dtype}"
+    if bias.dtype != weight.dtype:
+        return f"bias dtype {bias.dtype} != weight dtype {weight.dtype}"
+    return None
+
+
+def can_fuse_linear_gelu_static(linear: Any) -> bool:
+    """Input-independent guards: whether ``linear`` may ever use the epilogue."""
+    return _static_reject_reason(linear) is None
+
+
+def can_fuse_linear_gelu(linear: Any, x: torch.Tensor) -> bool:
+    """Whether ``gelu(linear(x))`` can use the fused cublasLt epilogue now."""
+    if not (x.is_cuda and x.dtype in (torch.bfloat16, torch.float16)):
+        return False
+    if getattr(linear, "weight", None) is None or x.dtype != linear.weight.dtype:
+        return False
+    return can_fuse_linear_gelu_static(linear)
+
+
+def mark_fused_gelu_site(module: nn.Module, linear_attr: str) -> None:
+    """Declare ``module`` as a tanh-GELU up-projection fusion site.
+
+    ``getattr(module, linear_attr)`` must be the up-projection linear whose
+    output feeds ``F.gelu(..., approximate="tanh")``. The site starts unmounted
+    (``_sgl_fused_gelu_enabled = False``): the module's forward must keep the
+    reference path bit-exact until :func:`mount_fused_linear_gelu` enables it.
+    """
+    setattr(module, _SITE_LINEAR_ATTR, linear_attr)
+    setattr(module, _SITE_ENABLED_ATTR, False)
+
+
+def iter_fused_gelu_sites(root: nn.Module) -> Iterator[nn.Module]:
+    """Yield every marked fusion site under ``root`` (including ``root``)."""
+    for module in root.modules():
+        if getattr(module, _SITE_LINEAR_ATTR, None) is not None:
+            yield module
+
+
+def mount_fused_linear_gelu(root: nn.Module) -> bool:
+    """Enable the fused epilogue on every marked site under ``root``.
+
+    All-or-nothing: if any marked site fails the static guards, every site is
+    left (or reset) on the reference path and False is returned. Returns False
+    as well when ``root`` has no marked sites.
+    """
+    sites = list(iter_fused_gelu_sites(root))
+    if not sites:
+        return False
+    for site in sites:
+        linear = getattr(site, getattr(site, _SITE_LINEAR_ATTR), None)
+        reason = "missing linear" if linear is None else _static_reject_reason(linear)
+        if reason is not None:
+            unmount_fused_linear_gelu(root)
+            logger.info(
+                "fused linear+GELU: %s site failed static guards (%s); "
+                "keeping the whole model on the reference path",
+                type(site).__name__,
+                reason,
+            )
+            return False
+    for site in sites:
+        setattr(site, _SITE_ENABLED_ATTR, True)
+    return True
+
+
+def unmount_fused_linear_gelu(root: nn.Module) -> None:
+    """Reset every marked site under ``root`` to the bit-exact reference path."""
+    for site in iter_fused_gelu_sites(root):
+        setattr(site, _SITE_ENABLED_ATTR, False)

python/sglang/multimodal_gen/runtime/models/dits/glm_image.py

@@ -18,6 +18,11 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 
+from sglang.kernels.ops.diffusion.fused_linear_gelu import (
+    can_fuse_linear_gelu,
+    fused_linear_gelu_tanh,
+    mark_fused_gelu_site,
+)
 from sglang.multimodal_gen.configs.models.dits.glmimage import GlmImageDitConfig
 from sglang.multimodal_gen.runtime.distributed.parallel_state import (
     get_sp_parallel_rank,
@@ -330,8 +335,17 @@ class GlmImageGELU(nn.Module):
             quant_config=quant_config,
             prefix=f"{prefix}.proj" if prefix else "proj",
         )
+        # quality="high" fusion site: up-proj GEMM + tanh-GELU in the cublasLt
+        # epilogue. Off by default; mounted per batch by the denoising stage.
+        mark_fused_gelu_site(self, "proj")
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self._sgl_fused_gelu_enabled and can_fuse_linear_gelu(
+            self.proj, hidden_states
+        ):
+            return fused_linear_gelu_tanh(
+                hidden_states, self.proj.weight, self.proj.bias
+            )
         hidden_states, _ = self.proj(hidden_states)
         return F.gelu(hidden_states, approximate="tanh")
 

python/sglang/multimodal_gen/runtime/models/dits/qwen_image.py

@@ -14,6 +14,11 @@ from diffusers.models.embeddings import TimestepEmbedding, Timesteps
 from diffusers.models.modeling_outputs import Transformer2DModelOutput
 from diffusers.models.normalization import AdaLayerNormContinuous
 
+from sglang.kernels.ops.diffusion.fused_linear_gelu import (
+    can_fuse_linear_gelu,
+    fused_linear_gelu_tanh,
+    mark_fused_gelu_site,
+)
 from sglang.multimodal_gen.configs.models.dits.qwenimage import QwenImageDitConfig
 from sglang.multimodal_gen.runtime.distributed import (
     get_local_torch_device,
@@ -851,8 +856,17 @@ class QwenImageGELU(nn.Module):
                 quant_config=quant_config,
                 prefix=f"{prefix}.proj",
             )
+        # quality="high" fusion site: up-proj GEMM + tanh-GELU in the cublasLt
+        # epilogue. Off by default; mounted per batch by the denoising stage.
+        mark_fused_gelu_site(self, "proj")
 
     def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        if self._sgl_fused_gelu_enabled and can_fuse_linear_gelu(
+            self.proj, hidden_states
+        ):
+            return fused_linear_gelu_tanh(
+                hidden_states, self.proj.weight, self.proj.bias
+            )
         hidden_states, _ = self.proj(hidden_states)
         return F.gelu(hidden_states, approximate="tanh")
 

python/sglang/multimodal_gen/runtime/pipelines_core/stages/denoising.py

@@ -20,6 +20,10 @@ from typing import Any
 import torch
 import torch.nn as nn
 
+from sglang.kernels.ops.diffusion.fused_linear_gelu import (
+    mount_fused_linear_gelu,
+    unmount_fused_linear_gelu,
+)
 from sglang.multimodal_gen import envs
 from sglang.multimodal_gen.configs.pipeline_configs.base import ModelTaskType, STA_Mode
 from sglang.multimodal_gen.configs.pipeline_configs.flux import (
@@ -222,6 +226,9 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
         # cache-dit state (for delayed mounting and idempotent control)
         self._cache_dit_enabled = False
         self._cached_num_steps = None
+        # fused linear+GELU state: whether the cublasLt-epilogue fusion is
+        # currently mounted on the transformers (quality="high" batches only).
+        self._fused_gelu_mounted = False
         self._torch_compile_registry = CompiledModuleRegistry()
         # Breakable CUDA graph runners, one per transformer module (lazy).
         self._bcg_runners: dict[int, Any] = {}
@@ -443,10 +450,39 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
         self, num_inference_steps: int | tuple[int, int], batch: Req
     ) -> None:
         """Apply request-dependent transformer acceleration in trace-safe order."""
+        self._maybe_toggle_fused_gelu(batch)
         self._maybe_enable_cache_dit(num_inference_steps, batch)
         for transformer in filter(None, [self.transformer, self.transformer_2]):
             self._maybe_torch_compile(transformer)
 
+    def _maybe_toggle_fused_gelu(self, batch: Req) -> None:
+        """Mount/unmount the cublasLt linear+GELU fusion for this batch.
+
+        The fused epilogue is numerically equivalent only at half-precision
+        rounding level (not bit-exact), so it is mounted for
+        ``quality="high"`` requests and unmounted otherwise -- the
+        ``"lossless"`` default runs the unmodified reference path bit-for-bit.
+        ``quality`` participates in the dynamic-batch signature, so a worker
+        batch is uniform in ``quality`` and this process-wide transition is
+        safe at the batch boundary. Mounting is all-or-nothing per
+        transformer (any ineligible marked site keeps the whole transformer
+        on the reference path); models without marked sites are no-ops.
+        """
+        want = getattr(batch.sampling_params, "quality", "lossless") == "high"
+        if want == self._fused_gelu_mounted:
+            return
+        mounted = False
+        for transformer in filter(None, [self.transformer, self.transformer_2]):
+            if want:
+                mounted |= mount_fused_linear_gelu(transformer)
+            else:
+                unmount_fused_linear_gelu(transformer)
+        self._fused_gelu_mounted = want
+        if want and mounted:
+            logger.info(
+                "Mounted fused linear+GELU (cublasLt epilogue) for quality=high"
+            )
+
     def _cache_dit_dual_model_name(self) -> str:
         return "wan2.2"