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 16, 04:30 UTC
[Bugfix] Use runner_type instead of task in GritLM (#11144)
Signed-off-by: Pooya Davoodi <pooya.davoodi@parasail.io>
Pooya Davoodi committed
Dec 12, 2024 at 20:09 UTC
1efce686053c15cd6f84361bb0bd1898fbb23a82
2 files changed
+7
-7
tests/models/embedding/language/test_gritlm.py
+3
-3
@@ -35,7 +35,7 @@ def test_find_array(monkeypatch):
35
from vllm.model_executor.models.gritlm import GritLMPooler
36
37
# Create an LLM object to get the model config.
38
- llm = vllm.LLM(MODEL_NAME, task="embedding", max_model_len=MAX_MODEL_LEN)
38
+ llm = vllm.LLM(MODEL_NAME, task="embed", max_model_len=MAX_MODEL_LEN)
39
pooler = GritLMPooler(model_config=llm.llm_engine.model_config)
40
41
arr = _arr([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
@@ -55,7 +55,7 @@ def server_embedding():
55
with pytest.MonkeyPatch.context() as mp:
56
mp.setenv("VLLM_ATTENTION_BACKEND", "XFORMERS")
57
58
- args = ["--task", "embedding", "--max_model_len", str(MAX_MODEL_LEN)]
58
+ args = ["--task", "embed", "--max_model_len", str(MAX_MODEL_LEN)]
59
with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
60
yield remote_server
61
@@ -141,7 +141,7 @@ def test_gritlm_offline_embedding(monkeypatch):
141
142
queries, q_instruction, documents, d_instruction = get_test_data()
143
144
- llm = vllm.LLM(MODEL_NAME, task="embedding", max_model_len=MAX_MODEL_LEN)
144
+ llm = vllm.LLM(MODEL_NAME, task="embed", max_model_len=MAX_MODEL_LEN)
145
146
d_rep = run_llm_encode(
147
llm,
vllm/model_executor/models/gritlm.py
+4
-4
@@ -203,12 +203,12 @@ class GritLM(LlamaForCausalLM):
203
) -> None:
204
super().__init__(vllm_config=vllm_config, prefix=prefix, **kwargs)
205
206
- self.task = vllm_config.model_config.task
206
+ self.runner_type = vllm_config.model_config.runner_type
207
208
self._pooler = GritLMPooler(vllm_config.model_config)
209
210
for layer in self.model.layers:
211
- if self.task == "embedding" and hasattr(layer, "self_attn"):
211
+ if self.runner_type == "pooling" and hasattr(layer, "self_attn"):
212
assert isinstance(layer.self_attn.attn.impl, XFormersImpl), (
213
"GritLM embedding is only supported by XFormers backend, "
214
"which can be forced by VLLM_ATTENTION_BACKEND=XFORMERS")
@@ -222,8 +222,8 @@ class GritLM(LlamaForCausalLM):
222
**kwargs,
223
) -> Union[torch.Tensor, IntermediateTensors]:
224
225
- # Change attention to non-causal for embedding task.
226
- if self.task == "embedding":
225
+ # Change attention to non-causal for pooling tasks.
226
+ if self.runner_type == "pooling":
227
assert attn_metadata.prefill_metadata.attn_bias is None
228
attn_metadata.prefill_metadata.attn_bias = [
229
BlockDiagonalMask.from_seqlens(attn_metadata.seq_lens)