biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Hardware][PPC64LE] Enable V1 for ppc64le and ARM (#20554)

Browse Source 
Signed-off-by: Akash Kaothalkar <akash.kaothalkar@ibm.com>
Co-authored-by: Akash Kaothalkar <akash.kaothalkar@ibm.com>
Co-authored-by: Nikhil Gupta <nikhil.gupta2@arm.com>
This commit is contained in:
Akash kaothalkar
2025-07-09 08:30:41 +05:30
committed by GitHub
parent 977180c912
commit 6db31e7a27
4 changed files with 77 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/v1/attention/backends/cpu_attn.py
View File

@@ -316,7 +316,6 @@ class TorchSDPAMetadataBuilderV1(AttentionMetadataBuilder[TorchSDPAMetadata]):
block_table: BlockTable) -> None:
self.runner = runner
self.block_table = block_table
# For reorder
self.reorder_prompt_req_index_list = np.empty(self.runner.max_num_reqs,
dtype=np.int64)
@@ -401,11 +400,14 @@ class TorchSDPAMetadataBuilderV1(AttentionMetadataBuilder[TorchSDPAMetadata]):
num_prefill_tokens=num_prefill_tokens,
num_decode_tokens=num_decode_tokens,
slot_mapping=slot_mapping,
# to ensure inference when chunked_prefill is disabled
seq_lens=runner.seq_lens_cpu[:num_reqs].tolist(),
seq_lens_tensor=runner.
seq_lens_cpu[num_prompt_req:num_reqs], # decode
max_decode_seq_len=max_decode_seq_len, # decode
block_tables=block_table_tensor[num_prompt_req:num_reqs], # decode
chunked_prefill=True,
chunked_prefill=self.runner.scheduler_config.
chunked_prefill_enabled,
max_query_len=max_query_len,
max_kv_len=max_prefill_seq_len,
prefill_query_start_loc=runner.