update self.kv_cache_dtype

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

@@ -1308,7 +1308,7 @@ class MLACommonImpl(MLAAttentionImpl[M], Generic[M]):
             decode_q_nope = decode_q_nope.transpose(0, 1)
             # Multiply (N, B, P) x (N, P, L) -> (N, B, L)
             # todo: bmm support
-            decode_ql_nope = torch.bmm(q_scale, decode_q_nope, self.W_UK_T) if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" and kv_cache_dtype_str=="fp8_e4m3" and envs.VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA else torch.bmm(decode_q_nope, self.W_UK_T) 
+            decode_ql_nope = torch.bmm(q_scale, decode_q_nope, self.W_UK_T) if torch.cuda.get_device_properties("cuda").gcnArchName.split(':')[0] == "gfx938" and self.kv_cache_dtype=="fp8_e4m3" and envs.VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA else torch.bmm(decode_q_nope, self.W_UK_T) 
             # Convert from (N, B, L) to (B, N, L)
             decode_ql_nope = decode_ql_nope.transpose(0, 1)