update deepgemm interface

vllm/model_executor/models/deepseek_v2.py

@@ -637,8 +637,8 @@ def sparse_attn_indexer(
                 )
             else:
                 logits = op.mqa_logits(
-                q_fp8[chunk.token_start:chunk.token_end],
-                (k_fp8, k_scale),
+                q_fp8[chunk.token_start:chunk.token_end].half(),
+                (k_fp8.half(), k_scale),
                 weights[chunk.token_start:chunk.token_end],
                 chunk.cu_seqlen_ks,
                 chunk.cu_seqlen_ke,
@@ -691,11 +691,9 @@ def sparse_attn_indexer(
                 max_model_len=max_model_len,
             )
         else:
-            padded_q_fp8_decode_tokens = padded_q_fp8_decode_tokens.half
-            kv_cache = kv_cache.half
             logits = gemmopt.paged_mqa_logits(
-                padded_q_fp8_decode_tokens,
-                kv_cache,
+                padded_q_fp8_decode_tokens.half(),
+                kv_cache.half(),
                 weights[:num_padded_tokens],
                 decode_metadata.seq_lens,
                 decode_metadata.block_table,