[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
@@ -38,7 +38,10 @@ from vllm.model_executor.layers.quantization.base_config import (
QuantizationConfig,
QuantizeMethodBase,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped
from vllm.model_executor.layers.quantization.utils.quant_utils import (
QuantKey,
is_layer_skipped,
)
from vllm.model_executor.utils import (
is_weights_pre_processed,
replace_parameter,
@@ -541,9 +544,12 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
supports_pre_processed_weights = True
def __init__(self, moe: FusedMoEConfig):
def __init__(self, moe: FusedMoEConfig, activation_key: QuantKey | None = None):
super().__init__(moe)
self.weight_dtype = "mxfp4"
# Checkpoint-declared activation quantization. Only backends that pick
# their activation format from the layer (Humming) consume it.
self.activation_key = activation_key
self.is_k3_situ_aiter = _use_k3_situ_aiter(moe)
self.is_k3_situ_int4_gfx942 = _use_k3_situ_int4_gfx942(moe)
self.experts_cls: type[mk.FusedMoEExperts] | None
@@ -822,6 +828,7 @@ class Mxfp4MoEMethod(FusedMoEMethodBase):
_cache_permute_indices=self._cache_permute_indices,
activation=self.moe.activation,
use_separated_a4w4=self.moe.use_mxfp4_w4a4_dsv4,
activation_key=self.activation_key,
)
)