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

[v1] Add Whisper model support (encoder-decoder) (#21088)

Browse Source 
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Co-authored-by: NickLucche <nlucches@redhat.com>
This commit is contained in:
Russell Bryant
2025-09-10 16:53:35 -04:00
committed by GitHub
parent 4db4426404
commit 37e8182bfe
31 changed files with 429 additions and 92 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm/v1/attention/backends/mamba_attn.py
View File

@@ -22,12 +22,9 @@ class BaseMambaAttentionMetadataBuilder(AttentionMetadataBuilder[M], abc.ABC):
def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
vllm_config: VllmConfig, device: torch.device):
assert isinstance(kv_cache_spec, MambaSpec)
self.kv_cache_spec = kv_cache_spec
self.device = device
self.vllm_config = vllm_config
self.layer_names = layer_names
super().__init__(kv_cache_spec, layer_names, vllm_config, device)
assert isinstance(kv_cache_spec, MambaSpec)
self.compilation_config = vllm_config.compilation_config
self.decode_cudagraph_max_bs = min(
self.vllm_config.scheduler_config.max_num_seqs,
@@ -52,4 +49,4 @@ class BaseMambaAttentionMetadataBuilder(AttentionMetadataBuilder[M], abc.ABC):
m.max_query_len = 1 # decode-only
return self.build(0, m)
return self.build(0, m)