[gpt-oss] Add gpt-oss mxfp4 support

vllm/v1/attention/backends/pallas.py

@@ -110,7 +110,6 @@ class PallasAttentionBackendImpl(AttentionImpl):
         alibi_slopes: Optional[list[float]],
         sliding_window: Optional[int],
         kv_cache_dtype: str,
-        blocksparse_params: Optional[dict[str, Any]] = None,
         logits_soft_cap: Optional[float] = None,
         attn_type: str = AttentionType.DECODER,
         kv_sharing_target_layer_name: Optional[int] = None,
@@ -120,9 +119,6 @@ class PallasAttentionBackendImpl(AttentionImpl):
             logger.warning_once(
                 "Using irope in Pallas is not supported yet, it will fall back "
                 "to global attention for long context.")
-        if blocksparse_params is not None:
-            raise ValueError("Paged attention Pallas kernel does "
-                             "not support block-sparse attention.")
         self.num_heads = num_heads
         self.head_size = head_size
         self.scale = float(scale)
@@ -139,8 +135,6 @@ class PallasAttentionBackendImpl(AttentionImpl):
             raise NotImplementedError("Alibi slopes is not supported.")
         if kv_cache_dtype != "auto":
             raise NotImplementedError("FP8 KV cache dtype is not supported.")
-        if blocksparse_params is not None:
-            raise NotImplementedError("Blocksparse is not supported.")
 
         if attn_type != AttentionType.DECODER:
             raise NotImplementedError("Encoder self-attention and "
@@ -161,6 +155,7 @@ class PallasAttentionBackendImpl(AttentionImpl):
         kv_cache: torch.Tensor,
         attn_metadata: PallasMetadata,
         output: Optional[torch.Tensor] = None,
+        output_scale: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
         """Forward pass with Pallas attention.
 
@@ -173,6 +168,11 @@ class PallasAttentionBackendImpl(AttentionImpl):
         Returns:
             shape = [num_tokens, num_heads * head_size]
         """
+        if output_scale is not None:
+            raise NotImplementedError(
+                "fused output quantization is not yet supported"
+                " for PallasAttentionBackendImpl")
+
         # For determine_available_memory case.
         if kv_cache.numel() == 0:
             if output is None: