remove conch-triton-kernels and skip AiterFlashAttentionMetadata

requirements/rocm.txt

@@ -17,4 +17,4 @@ setuptools>=77.0.3,<80.0.0
 setuptools-scm>=8
 runai-model-streamer==0.11.0
 runai-model-streamer-s3==0.11.0
-conch-triton-kernels==1.2.1
+# conch-triton-kernels==1.2.1

vllm/v1/spec_decode/eagle.py

@@ -20,8 +20,8 @@ from vllm.model_executor.models.llama_eagle3 import Eagle3LlamaForCausalLM
 from vllm.platforms import current_platform
 from vllm.utils import is_pin_memory_available
 from vllm.v1.attention.backends.flash_attn import FlashAttentionMetadata
-from vllm.v1.attention.backends.rocm_aiter_fa import (
-    AiterFlashAttentionMetadata)
+# from vllm.v1.attention.backends.rocm_aiter_fa import (
+#     AiterFlashAttentionMetadata)
 from vllm.v1.attention.backends.tree_attn import (TreeAttentionMetadata,
                                                   TreeAttentionMetadataBuilder)
 from vllm.v1.attention.backends.triton_attn import TritonAttentionMetadata
@@ -231,7 +231,7 @@ class EagleProposer:
         if current_platform.is_rocm():
             assert isinstance(
                 attn_metadata,
-                (TritonAttentionMetadata, AiterFlashAttentionMetadata,
+                (TritonAttentionMetadata, # AiterFlashAttentionMetadata,
                  FlashAttentionMetadata))
         else:
             # Currently, only FlashAttention supports multi-token eagle spec