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

Remove V0 Encoder-Decoder Support (#24907)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk@thinkingmachines.ai>
This commit is contained in:
Woosuk Kwon
2025-09-15 21:17:14 -07:00
committed by GitHub
parent 5206ab20ba
commit 759ef49b15
47 changed files with 13 additions and 9661 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
tests/utils_/test_utils.py
View File

@@ -501,34 +501,6 @@ def test_bind_kv_cache_non_attention():
assert ctx['model.layers.28.attn'].kv_cache[0] is kv_cache[1]
def test_bind_kv_cache_encoder_decoder(monkeypatch: pytest.MonkeyPatch):
# V1 TESTS: ENCODER_DECODER is not supported on V1 yet.
with monkeypatch.context() as m:
m.setenv("VLLM_USE_V1", "0")
from vllm.attention import Attention, AttentionType
# example from bart
ctx = {
'encoder.layers.0.self_attn.attn':
Attention(32, 128, 0.1, attn_type=AttentionType.ENCODER),
'decoder.layers.0.encoder_attn.attn':
Attention(32, 128, 0.1, attn_type=AttentionType.ENCODER_DECODER),
'decoder.layers.0.self_attn.attn':
Attention(32, 128, 0.1, attn_type=AttentionType.DECODER),
}
kv_cache = [
torch.zeros((1, )),
]
encoder_kv_cache = ctx['encoder.layers.0.self_attn.attn'].kv_cache
bind_kv_cache(ctx, [kv_cache])
assert ctx['encoder.layers.0.self_attn.attn'].kv_cache is encoder_kv_cache
assert ctx['decoder.layers.0.encoder_attn.attn'].kv_cache[0] is kv_cache[0]
assert ctx['decoder.layers.0.self_attn.attn'].kv_cache[0] is kv_cache[0]
def test_bind_kv_cache_pp():
with patch("vllm.utils.cuda_device_count_stateless", lambda: 2):
# this test runs with 1 GPU, but we simulate 2 GPUs