[ROCm][Bugfix] Revert #56433 + #51692 to fix accuracy breakdown for DeepSeek-V4 (#57132)

Signed-off-by: Shanshan Shen <87969357+shen-shanshan@users.noreply.github.com>
This commit is contained in:
Shanshan Shen
2026-09-16 18:58:23 +00:00
committed by GitHub
parent 403f182c02
commit bc0f47cd03
8 changed files with 34 additions and 402 deletions
+6 -11
View File
@@ -71,10 +71,7 @@ from vllm.model_executor.models.utils import (
make_layers,
maybe_prefix,
)
from vllm.models.deepseek_v4.amd.rocm import (
DeepseekV4ROCMAiterMLAAttention,
weight_already_preshuffled,
)
from vllm.models.deepseek_v4.amd.rocm import DeepseekV4ROCMAiterMLAAttention
from vllm.platforms import current_platform
from vllm.platforms.rocm import on_gfx950
from vllm.sequence import IntermediateTensors
@@ -152,13 +149,11 @@ class DeepseekV4MLP(nn.Module):
return
if ws.dtype == torch.float8_e8m0fnu:
ws = _upcast_e8m0_to_fp32(ws).contiguous()
# Skip if the linear's kernel already shuffled it.
if not weight_already_preshuffled(self.gate_up_proj):
replace_parameter(
self.gate_up_proj,
"weight",
rocm_aiter_ops.shuffle_weight(w.data, layout=(16, 16)),
)
replace_parameter(
self.gate_up_proj,
"weight",
rocm_aiter_ops.shuffle_weight(w.data, layout=(16, 16)),
)
self._gateup_scale = ws
def forward(self, x):