[diffusion][feature] Add offline SVDQuant W4A4 support (#6162)

Signed-off-by: lishunyang12 <lishunyang12@163.com>
Co-authored-by: Hongsheng Liu <liuhongsheng4@huawei.com>
This commit is contained in:
SYLARandHongsheng Liu authored and GitHub committed 2026-08-27 08:45:53 +08:00
1 parent 432b2c6ca3
commit 0e44fc7406
9 files changed
+932 -20

No files matched your search

+6 -5
View File
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
"""Per-component quantization routing for multi-stage models.
Routes get_quant_method() to different configs based on longest-prefix match:
@@ -26,10 +26,11 @@ if TYPE_CHECKING:
)
# Pre-quantized checkpoints (modelopt FP8/FP4/MXFP8/mixed) only quantize the
# Thinker LM. Vision and audio encoder weights remain in BF16 with no
# corresponding scale tensors in the checkpoint.
PRE_QUANTIZED_METHODS: frozenset[str] = frozenset({"modelopt", "modelopt_fp4", "modelopt_mxfp8", "modelopt_mixed"})
# These pre-quantized formats require serialized scale or correction tensors
# that the vision and audio encoder checkpoints do not provide.
PRE_QUANTIZED_METHODS: frozenset[str] = frozenset(
{"modelopt", "modelopt_fp4", "modelopt_mxfp8", "modelopt_mixed", "svdquant"}
)
def resolve_component_quant_config(