Merge branch 'v0.15.1-mqa_fp8' into 'v0.15.1-dev'

feat:支持mqa的fp8实现

See merge request dcutoolkit/deeplearing/vllm!514

vllm/model_executor/layers/sparse_attn_indexer.py

@@ -112,7 +112,7 @@ def sparse_attn_indexer(
                     chunk.cu_seqlen_ks,
                     chunk.cu_seqlen_ke,
                 )
-            else:
+            elif torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938":
                 k_fp8 = k_fp8_full[: chunk.total_seq_lens]
                 k_scale = k_scale_full[: chunk.total_seq_lens]
                 ops.cp_gather_indexer_k_quant_cache(
@@ -121,7 +121,21 @@ def sparse_attn_indexer(
                     k_scale,
                     chunk.block_table,
                     chunk.cu_seq_lens,
-                )                
+                )
+                logits = op.mqa_logits(
+                    q_fp8[chunk.token_start:chunk.token_end],  
+                    k_fp8, 
+                    weights[chunk.token_start:chunk.token_end], 
+                    chunk.cu_seqlen_ks, 
+                    chunk.cu_seqlen_ke,
+                    q_fp8[chunk.token_start:chunk.token_end].shape[0],
+                    k_fp8.shape[0],
+                    q_fp8.shape[1],
+                    q_fp8.shape[2],
+                    k_scale.view(torch.float32).flatten(),
+                    True
+                    )
+            else:                
                 logits = op.mqa_logits(
                     q_fp8[chunk.token_start:chunk.token_end],  
                     k,