@tduncan / Nvidia-Optimization-Stack-VLLM Mirror · read-only / commits / ea5f14e6ff
Fork 0

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 17, 04:15 UTC

[Bugfix][Model] Fix Qwen2 (#3554)

Roy committed Mar 22, 2024 at 08:18 UTC ea5f14e6ffafcb9c660a3eea5a935122aa9f84ae
2 files changed +2 -1
tests/models/test_models.py
+1
@@ -20,6 +20,7 @@ MODELS = [
20 "stabilityai/stablelm-3b-4e1t",
21 "allenai/OLMo-1B",
22 "bigcode/starcoder2-3b",
23 + "Qwen/Qwen1.5-0.5B",
24 ]
25
26
vllm/model_executor/models/qwen2.py
+1 -1
@@ -349,7 +349,7 @@ class Qwen2ForCausalLM(nn.Module):
349 ("gate_up_proj", "gate_proj", 0),
350 ("gate_up_proj", "up_proj", 1),
351 ]
352 - params_dict = dict(self.named_parameters())
352 + params_dict = dict(self.named_parameters(remove_duplicate=False))
353 for name, loaded_weight in hf_model_weights_iterator(
354 model_name_or_path, cache_dir, load_format, revision):
355 if "rotary_emb.inv_freq" in name: