chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Fix the run-time error when calling fused_rms_mxfp4_quant that change return output number (#12803)

Browse Source 
Co-authored-by: wunhuang <wunhuang@amd.com>
This commit is contained in:
kk
2025-11-10 13:48:36 +08:00
committed by GitHub
parent 9cfe78dd30
commit 90401cf7d2
3 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
python/sglang/srt/layers/communicator.py
View File

@@ -268,7 +268,7 @@ class LayerCommunicator:
residual = hidden_states
if _use_aiter and _is_gfx95_supported and ("mxfp4" in qaunt_format):
hidden_states = fused_rms_mxfp4_quant(
hidden_states, *_, _ = fused_rms_mxfp4_quant(
hidden_states,
self.input_layernorm.weight,
self.input_layernorm.variance_epsilon,
@@ -281,7 +281,7 @@ class LayerCommunicator:
hidden_states = self.input_layernorm(hidden_states)
else:
if _use_aiter and _is_gfx95_supported and ("mxfp4" in qaunt_format):
hidden_states, residual = fused_rms_mxfp4_quant(
hidden_states, *_, residual = fused_rms_mxfp4_quant(
hidden_states,
self.input_layernorm.weight,
self.input_layernorm.variance_epsilon,

4
python/sglang/srt/models/deepseek_v2.py
View File

@@ -1608,7 +1608,7 @@ class DeepseekV2AttentionMLA(nn.Module):
current_stream.wait_stream(self.alt_stream)
else:
if _use_aiter_gfx95 and self.q_b_proj.weight.dtype == torch.uint8:
q, k_nope = fused_rms_mxfp4_quant(
q, k_nope, *_ = fused_rms_mxfp4_quant(
q,
self.q_a_layernorm.weight,
self.q_a_layernorm.variance_epsilon,
@@ -1937,7 +1937,7 @@ class DeepseekV2AttentionMLA(nn.Module):
current_stream.wait_stream(self.alt_stream)
else:
if _use_aiter_gfx95 and self.q_b_proj.weight.dtype == torch.uint8:
q, k_nope = fused_rms_mxfp4_quant(
q, k_nope, *_ = fused_rms_mxfp4_quant(
q,
self.q_a_layernorm.weight,
self.q_a_layernorm.variance_epsilon,