[diffusion] Z-Image: autotune the shared transformer block at both token scales (#35755)

This commit is contained in:
Frederik Gossen
2026-10-10 13:04:43 +08:00
committed by GitHub
parent aa79087017
commit 3831e7e091
4 changed files with 31 additions and 1 deletions
@@ -46,6 +46,10 @@ class DiTConfig(ModelConfig):
quant_config: QuantizationConfig | None = None
torch_compile_mode: str = "max-autotune-no-cudagraphs"
# torch._inductor.config overrides for this model, applied before compiling,
# e.g. {"multi_kernel_hints": [64, 4096]}.
torch_compile_inductor_config: dict[str, Any] = field(default_factory=dict)
@staticmethod
def add_cli_args(parser: Any, prefix: str = "dit-config") -> Any:
"""Add CLI arguments for DiTConfig fields"""
@@ -2,7 +2,7 @@
# SPDX-License-Identifier: Apache-2.0
from dataclasses import dataclass, field
from typing import Tuple
from typing import Any, Tuple
from sglang.multimodal_gen.configs.models.dits.base import DiTArchConfig, DiTConfig
@@ -96,3 +96,10 @@ class ZImageDitConfig(DiTConfig):
arch_config: ZImageArchConfig = field(default_factory=ZImageArchConfig)
prefix: str = "zimage"
# One block serves both the caption tokens (tens) and the image and unified
# streams (thousands), with a dynamic token dim, so its GEMMs get autotuned
# for one branch and run for all three -- 1.4x slower when that goes wrong.
torch_compile_inductor_config: dict[str, Any] = field(
default_factory=lambda: {"multi_kernel_hints": [64, 4096]}
)
@@ -173,6 +173,7 @@ from sglang.multimodal_gen.runtime.utils.precision import (
from sglang.multimodal_gen.runtime.utils.profiler import SGLDiffusionProfiler
from sglang.multimodal_gen.runtime.utils.torch_compile import (
CompiledModuleRegistry,
apply_inductor_config,
resolve_torch_compile_kwargs,
)
@@ -584,6 +585,10 @@ class DenoisingStage(PipelineStage, RolloutDenoisingMixin):
return
dit_config = getattr(self.server_args.pipeline_config, "dit_config", None)
if not current_platform.is_npu():
apply_inductor_config(
getattr(dit_config, "torch_compile_inductor_config", {})
)
compile_kwargs, mode = resolve_torch_compile_kwargs(
"SGLANG_TORCH_COMPILE_MODE",
config=dit_config,
@@ -8,6 +8,9 @@ from typing import Any
import torch.nn as nn
from sglang.multimodal_gen.runtime.platforms import current_platform
from sglang.multimodal_gen.runtime.utils.logging_utils import init_logger
logger = init_logger(__name__)
def maybe_enable_inductor_compute_comm_overlap() -> None:
@@ -19,6 +22,17 @@ def maybe_enable_inductor_compute_comm_overlap() -> None:
pass
def apply_inductor_config(overrides: dict[str, Any]) -> None:
"""Apply a model's torch._inductor.config overrides before it is compiled."""
import torch._inductor.config as _inductor_cfg
for name, value in overrides.items():
if not hasattr(_inductor_cfg, name):
logger.warning("Unknown torch._inductor.config option %r, ignoring", name)
continue
setattr(_inductor_cfg, name, value)
def build_torch_compile_kwargs(
*,
mode: str | None,