A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 6, 09:30 UTC
[Refactor] Remove dead code in pooling model (#37572)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
Wentao Ye committed
Mar 19, 2026 at 19:39 UTC
2be1a0f74b016a589c6392670b66a2c8413f1a6a
2 files changed
-12
vllm/entrypoints/pooling/utils.py
-8
@@ -60,14 +60,6 @@ def encode_pooling_output_float(output: PoolingRequestOutput) -> list[float]:
60
return output.outputs.data.tolist()
61
62
63
-def encode_pooling_output_binary(
64
- output: PoolingRequestOutput,
65
- embed_dtype: EmbedDType,
66
- endianness: Endianness,
67
-) -> bytes:
68
- return tensor2binary(output.outputs.data, embed_dtype, endianness)
69
-
70
-
63
def encode_pooling_output_base64(
64
output: PoolingRequestOutput,
65
embed_dtype: EmbedDType,
vllm/v1/pool/metadata.py
-4
@@ -14,7 +14,6 @@ pin_memory = is_pin_memory_available()
14
15
@dataclass
16
class PoolingCursor:
17
- index: list[int]
17
first_token_indices_gpu: torch.Tensor
18
last_token_indices_gpu: torch.Tensor
19
prompt_lens_cpu: torch.Tensor
@@ -23,7 +22,6 @@ class PoolingCursor:
22
23
def __getitem__(self, indices: slice):
24
return PoolingCursor(
26
- index=self.index[indices],
25
first_token_indices_gpu=self.first_token_indices_gpu[indices],
26
last_token_indices_gpu=self.last_token_indices_gpu[indices],
27
prompt_lens_cpu=self.prompt_lens_cpu[indices],
@@ -108,7 +106,6 @@ class PoolingMetadata:
106
107
assert len(prompt_lens) == n_seq
108
111
- index = list(range(n_seq))
109
num_scheduled_tokens_cpu = torch.from_numpy(num_scheduled_tokens_np)
110
if query_start_loc_gpu is None:
111
cumsum = torch.zeros(
@@ -130,7 +127,6 @@ class PoolingMetadata:
127
)
128
cumsum = query_start_loc_gpu
129
self.pooling_cursor = PoolingCursor(
133
- index=index,
130
first_token_indices_gpu=cumsum[:n_seq],
131
last_token_indices_gpu=cumsum[1:] - 1,
132
prompt_lens_cpu=prompt_lens,