karylab_agents/vllm
1
0
Fork 0
forked from Karylab-cklius/vllm
Code Pull Requests 2 Actions 1 Packages Activity

[ROCm][Bugfix]: W4A4 MOE using emulation instead of AITER on MXFP4-supported hardware (#41175)

Browse Source 
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
This commit is contained in:
Rohan Potdar
2026-04-29 16:39:03 -05:00
committed by GitHub
parent b58669cb42
commit 944e138bcf
2 changed files with 4 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py
+2 -1
View File
@@ -252,7 +252,8 @@ def rocm_aiter_fused_experts(
else:
quant_method = QuantMethod.NO.value
# mxfp4: both w4a4 (quark) and w4a16 (oracle CK) use BLOCK_1X32
# mxfp4 i.e. w4a4, w4a16 uses BLOCK_1X32
# mxfp6 and mxfp8 are unsupported in AITER currently and use emulation instead
if quant_config.use_mxfp4_w4a4 or quant_config.use_mxfp4_w4a16:
quant_method = QuantMethod.BLOCK_1X32.value
# w8a8 block-scaled
vllm/model_executor/layers/quantization/quark/quark_moe.py
+2 -6
View File
@@ -1025,12 +1025,8 @@ class QuarkOCP_MX_MoEMethod(QuarkMoEMethod):
get_current_vllm_config().model_config.hf_config, "model_type", None
)
# TODO(aiter): extend once rocm_aiter_fused_experts gains dispatch
# for the other OCP MX schemes. Today its CK MoE kernel only has an
# entry for `w_mxfp4` (w4a16); mixed schemes like `w_mxfp4_a_mxfp6_*`
# fall through to QuantMethod.NO and raise "Unsupported kernel config
# for moe heuristic dispatch".
_AITER_NATIVE_OCP_MX_SCHEMES = ("w_mxfp4",)
# TODO: Remove once all OCP MX schemes use the kernel abstraction
_AITER_NATIVE_OCP_MX_SCHEMES = ("w_mxfp4", "w_mxfp4_a_mxfp4")
self.emulate = (
not current_platform.supports_mx()
or self.ocp_mx_scheme not in _AITER_NATIVE_OCP_MX_SCHEMES