update paged_mqa_logits

vllm/model_executor/models/deepseek_v2.py

@@ -691,6 +691,8 @@ 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,
@@ -698,7 +700,7 @@ def sparse_attn_indexer(
                 decode_metadata.seq_lens,
                 decode_metadata.block_table,
                 decode_metadata.schedule_metadata,
-                max_model_len=max_model_len,
+                max_context_len=max_model_len,
             )
         # padded query len
         current_device = padded_q_fp8_decode_tokens.device

vllm/v1/attention/backends/mla/indexer.py

@@ -312,7 +312,7 @@ class DeepseekV32IndexerMetadataBuilder(AttentionMetadataBuilder):
             seq_lens = common_attn_metadata.seq_lens[:num_decodes]
 
             if current_platform.is_rocm():
-                self.scheduler_metadata_buffer[:] = gemmopt.get_paged_mqa_logits_metadata(
+                self.scheduler_metadata_buffer = gemmopt.get_paged_mqa_logits_metadata(
                     seq_lens, self.kv_cache_spec.block_size, self.num_sms)
             else:
                 self.scheduler_metadata_buffer[:] = get_paged_mqa_logits_metadata(