[Bugfix][ROCm][CI] Restore the DeepSeek-V4 input GEMM override point (#51821)

Signed-off-by: Stefan Koncarevic <Stefan.Koncarevic@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
stefankoncarevic
2026-08-12 20:33:52 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent 5936bac7d7
commit b369f10d5c
+8 -1
View File
@@ -443,6 +443,13 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
# Inverse-RoPE + wo_a + wo_b output projection (platform-specific).
return self._o_proj(o, positions)
def _fused_wqa_wkv_gemm(self, hidden_states: torch.Tensor) -> torch.Tensor:
# Override point: the ROCm layer preshuffles this weight in place, so
# it cannot go through fused_wqa_wkv directly.
# MergedColumnParallelLinear returns (output, bias); bias is None.
qr_kv, _ = self.fused_wqa_wkv(hidden_states)
return qr_kv
def _run_parallel_input_projections(
self, hidden_states: torch.Tensor
) -> tuple[
@@ -494,7 +501,7 @@ class DeepseekV4Attention(nn.Module, AttentionLayerBase, ABC):
aux_fns[2] = indexer_compressor_kv_score
qr_kv, (kv_score, indexer_weights, indexer_kv_score) = execute_in_parallel(
lambda: self.fused_wqa_wkv(hidden_states)[0],
lambda: self._fused_wqa_wkv_gemm(hidden_states),
aux_fns,
self.ln_events[0],
self.ln_events[1:4],