[ROCm][Perf] Add opt-in a4w4 (FP4 activation) MoE for DeepSeek V4.1 on AITER (#58819)

Signed-off-by: fai <fangzhouai@gmail.com>
Signed-off-by: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com>
Co-authored-by: Cursor Agent (Claude Sonnet 5) <cursoragent@cursor.com>
This commit is contained in:
Fangzhou Ai
2026-09-30 15:07:44 +08:00
committed by GitHub
co-authored by Cursor Agent
parent e5f3d08d72
commit 77e8b02c8d
8 changed files with 204 additions and 11 deletions
@@ -6,6 +6,7 @@ from typing import Union
import torch
import vllm.envs as envs
from vllm.config import ParallelConfig, SchedulerConfig
from vllm.config.kernel import MoEBackend
from vllm.distributed import get_dp_group, get_pcp_group, get_tensor_model_parallel_rank
@@ -1247,6 +1248,12 @@ class FusedMoEParallelConfig:
)
# Model types validated for VLLM_ROCM_USE_AITER_MOE_A4W4_DSV4=1. This flag
# also changes the MXFP4 weight shuffle layout, so using it on other models
# can produce garbled output.
_AITER_MOE_A4W4_DSV4_VALIDATED_MODEL_TYPES = ("deepseek_v41", "deepseek_v41_text")
# Adapted from pplx-kernels tests/all_to_all_utils.py
@dataclass
class FusedMoEConfig:
@@ -1314,6 +1321,11 @@ class FusedMoEConfig:
tp_shard_with_padding: bool = False
rocm_aiter_fmoe_enabled: bool = False
aiter_fmoe_shared_expert_enabled: bool = False
# Whether to force MXFP4 (a4w4) MoE activations for DeepSeek V4.1 on
# ROCm/AITER. Opt-in via VLLM_ROCM_USE_AITER_MOE_A4W4_DSV4=1; rejected
# for any other model type. Resolved here, not in the forward path,
# because get_current_vllm_config() isn't set there.
use_mxfp4_w4a4_dsv4: bool = False
def __post_init__(self):
from vllm._aiter_ops import rocm_aiter_ops
@@ -1345,6 +1357,29 @@ class FusedMoEConfig:
rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
)
if self.rocm_aiter_fmoe_enabled and envs.VLLM_ROCM_USE_AITER_MOE_A4W4_DSV4:
from vllm.config import get_current_vllm_config_or_none
vllm_config = get_current_vllm_config_or_none()
model_type = (
getattr(vllm_config.model_config.hf_config, "model_type", None)
if vllm_config is not None
else None
)
if model_type not in _AITER_MOE_A4W4_DSV4_VALIDATED_MODEL_TYPES:
raise ValueError(
f"VLLM_ROCM_USE_AITER_MOE_A4W4_DSV4=1 only supports "
f"model_type in {_AITER_MOE_A4W4_DSV4_VALIDATED_MODEL_TYPES}, "
f"got {model_type!r}. Unset this env var for this model."
)
if not rocm_aiter_ops.fused_moe_supports_quant_dtype_a():
raise ValueError(
"VLLM_ROCM_USE_AITER_MOE_A4W4_DSV4=1 needs an AITER build "
"with fused_moe(quant_dtype_a=...) support "
"(ROCm/aiter#5439+). Upgrade AITER or unset this env var."
)
self.use_mxfp4_w4a4_dsv4 = True
if self.use_mori_kernels:
assert self.rocm_aiter_fmoe_enabled, (
"Mori needs to be used with aiter fused_moe for now."