update DeepseekV32IndexerCache

vllm/model_executor/models/deepseek_v2.py

@@ -702,7 +702,7 @@ def sparse_attn_indexer(
         else:
             logits = gemmopt.paged_mqa_logits(
                 padded_q_fp8_decode_tokens, 
-                kv_cache if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else kv_cache.to(torch.bfloat16), 
+                kv_cache, 
                 weights[:num_padded_tokens] if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else weights[:num_padded_tokens].to(torch.float32), 
                 decode_metadata.seq_lens, 
                 decode_metadata.block_table, 
@@ -829,8 +829,8 @@ class Indexer(nn.Module):
         #       per self.quant_block_size element
         self.k_cache = DeepseekV32IndexerCache(
             head_dim=self.head_dim +
-            self.head_dim // self.quant_block_size * 4,
-            dtype=torch.uint8,
+            self.head_dim // self.quant_block_size * 4 if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else self.head_dim,
+            dtype=torch.uint8 if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" else torch.bfloat16,
             prefix=f"{prefix}.k_cache",
             cache_config=cache_config)
         self.max_model_len = vllm_config.model_config.max_model_len