Merge remote-tracking branch 'nm/sage/amd-deepseek' into upstream_merge_25_03_10

Author: gshtras

vllm/attention/backends/mla/common.py

@@ -1327,21 +1327,7 @@ def _compute_prefill_context(
                                                [0, q.shape[-1] - v.shape[-1]],
                                                value=0)
 
-            if is_hip and envs.VLLM_USE_TRITON_FLASH_ATTN:
-                attn_output, attn_softmax_lse = self.triton_fa_func(
-                    q,
-                    k,
-                    v_padded,
-                    None,
-                    prefill_metadata.query_start_loc,
-                    prefill_metadata.context_chunk_cu_seq_lens[i],
-                    prefill_metadata.max_query_len,
-                    prefill_metadata.context_chunk_max_seq_lens[i],
-                    False,  # causal
-                    self.scale,
-                    None,  # attn_mask is None unless applying ALiBi mask
-                )
-            elif is_vllm_fa:
+            if is_vllm_fa:
                 attn_output, attn_softmax_lse = self.flash_attn_varlen_func(
                     q=q,
                     k=k,

vllm/config.py

@@ -3474,9 +3474,9 @@ def __post_init__(self):
             self.compilation_config.level = CompilationLevel.NO_COMPILATION
 
         if self.model_config and self.model_config.use_mla and \
-            not current_platform.is_cuda():
+            not (current_platform.is_cuda() or current_platform.is_rocm()):
             logger.info(
-                "MLA is enabled on a non-cuda platform; forcing chunked "
+                "MLA is enabled on a non-GPU platform; forcing chunked "
                 "prefill and prefix caching to be disabled.")
             self.scheduler_config.enable_chunked_prefill = False
             self.scheduler_config.chunked_prefill_enabled = False