[Bugfix][Model] Fix FP8 PLE loading in mixed ModelOpt checkpoints (#54882)

Signed-off-by: Shiyang Chen <shiychen@nvidia.com>
Co-authored-by: Vadim Gimpelson <156319763+vadiklyutiy@users.noreply.github.com>
This commit is contained in:
sychen52
2026-09-03 21:09:39 +04:00
committed by GitHub
co-authored by Vadim Gimpelson
parent d410fc12f3
commit d4d703caf9
2 changed files with 42 additions and 0 deletions
@@ -22,6 +22,9 @@ from vllm.model_executor.layers.quantization.base_config import (
QuantizeMethodBase,
)
from vllm.model_executor.layers.quantization.fp8 import Fp8Config
from vllm.model_executor.layers.quantization.modelopt import (
ModelOptMixedPrecisionConfig,
)
from vllm.model_executor.layers.quantization.utils.fp8_utils import (
create_fp8_scale_parameter,
create_fp8_weight_parameter,
@@ -133,6 +136,11 @@ def _get_ple_embedding_quant_method(
) -> QuantizeMethodBase | None:
"""Select global-scale FP8 only for quantized PLE checkpoint shards."""
if isinstance(quant_config, ModelOptMixedPrecisionConfig):
if quant_config._resolve_quant_algo(prefix) == "FP8":
return Qwen4ExpPLEFp8EmbeddingMethod()
return None
if not isinstance(quant_config, Fp8Config):
return None
if not quant_config.is_checkpoint_fp8_serialized: