xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-03-28 10:39:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Bugfix][Hardware][Gaudi] Bump vllm_hpu_extension version (#11028)

Browse Source 
Signed-off-by: Konrad Zawora <kzawora@habana.ai>
This commit is contained in:
Konrad Zawora 2024-12-09 22:21:06 +01:00 committed by GitHub
parent a811dd6608
commit cbcbdb1ceb
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 12 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
requirements-hpu.txt
View File

@ -8,4 +8,4 @@ pandas
tabulate
setuptools>=61
setuptools-scm>=8
vllm-hpu-extension @ git+https://github.com/HabanaAI/vllm-hpu-extension.git@fd7f2e6
vllm-hpu-extension @ git+https://github.com/HabanaAI/vllm-hpu-extension.git@e096d6f

11
vllm/attention/backends/hpu_attn.py
View File

@ -111,8 +111,16 @@ class HPUAttentionImpl(AttentionImpl, torch.nn.Module):
self.matmul_qk = Matmul()
self.softmax = Softmax()
self.matmul_av = Matmul()
self.batch2block_matmul = Matmul()
self.block2batch_matmul = Matmul()
# NOTE(kzawora): Contiguous PA is off until model runner supports it
self.k_cache = VLLMKVCache()
self.k_cache.use_contiguous_pa = False
self.v_cache = VLLMKVCache()
self.v_cache.use_contiguous_pa = False
# NOTE(kzawora): Pipelined PA is off until model runner supports it
ops.pa_impl = ops.pa
self.num_kv_heads = num_heads if num_kv_heads is None else num_kv_heads
self.sliding_window = sliding_window
self.alibi_slopes = alibi_slopes
@ -228,9 +236,12 @@ class HPUAttentionImpl(AttentionImpl, torch.nn.Module):
block_mapping=attn_metadata.block_mapping,
block_bias=attn_metadata.attn_bias,
block_scales=attn_metadata.block_scales,
block_groups=None,
scale=self.scale,
matmul_qk_op=self.matmul_qk,
matmul_av_op=self.matmul_av,
batch2block_matmul_op=self.batch2block_matmul,
block2batch_matmul_op=self.block2batch_matmul,
keys_fetch_func=self.k_cache.fetch_from_cache,
values_fetch_func=self.v_cache.fetch_from_cache)
# Reshape the output tensor.