[Quantization] Support compressed-tensors MXFP4 MoE with FP8 activations via Mxfp4MoEMethod (#59815)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Misha Goin
2026-10-14 13:33:35 -07:00
committed by GitHub
co-authored by Claude Kevin H. Luu
parent fe98d90a9d
commit 984c242ede
12 changed files with 311 additions and 10 deletions
@@ -1480,6 +1480,7 @@ def convert_weight_to_mxfp4_moe_kernel_format(
_cache_permute_indices: dict[torch.Size, torch.Tensor] | None = None,
activation: MoEActivation | None = None,
use_separated_a4w4: bool = False,
activation_key: QuantKey | None = None,
) -> tuple[
torch.Tensor,
torch.Tensor,
@@ -1541,8 +1542,20 @@ def convert_weight_to_mxfp4_moe_kernel_format(
convert_to_humming_moe_kernel_format,
)
# The env override wins; otherwise use the checkpoint's activation
# scheme instead of Humming's default BF16 activations.
input_schema = None
if activation_key is not None and not envs.VLLM_HUMMING_INPUT_QUANT_CONFIG:
from vllm.model_executor.layers.quantization.utils.humming import (
quant_key_to_input_schema,
)
input_schema = quant_key_to_input_schema(activation_key)
convert_to_humming_moe_kernel_format(
layer, quant_config={"quant_method": "mxfp4"}
layer,
quant_config={"quant_method": "mxfp4"},
input_schema=input_schema,
allow_input_schema_fallback=False,
)
return (
layer.w13_weight,