[Hardware][Ascend] Add Wan2.2 MXFP4 UOS, Smooth and W4A8 fallback (#7210)

Signed-off-by: HAAZZZEEEE <wang-taicheng@qq.com>
This commit is contained in:
HAAZZZEEEE authored and GitHub committed 2026-09-17 03:11:40 +00:00
1 parent 4fab34c8f4
commit becfefd440
20 files changed
+2435 -184

No files matched your search

+4
View File
@@ -1021,6 +1021,8 @@ class OmniDiffusionConfig:
# str is resolved to {"method": <str>} internally.
# Per-component: {"transformer": {"method": "fp8"}, "vae": None}
quantization_config: str | QuantizationConfig | dict[str, Any] | None = None
# Internal provenance, retained across config projection and worker transport.
quantization_config_is_auto_detected: bool = False
# Explicit runtime override for ModelOpt FP8 diffusion checkpoints. This
# does not enable FP8 by itself; it only selects CUTLASS once the checkpoint
# has already resolved to vLLM's ModelOpt FP8 linear method.
@@ -1358,6 +1360,8 @@ class OmniDiffusionConfig:
or (is_checkpoint_nvfp4 and self._is_generic_nvfp4_quant_config(self.quantization_config))
)
if should_use_checkpoint_config:
if self.quantization_config is None:
self.quantization_config_is_auto_detected = True
self.quantization_config = tf_config.quant_config
logger.info(
"Auto-detected quantization '%s' from model config",