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

[Model] Support for fairseq2 Llama (#11442)

Browse Source 
Signed-off-by: Martin Gleize <mgleize@meta.com>
Co-authored-by: mgleize user <mgleize@a100-st-p4de24xlarge-4.fair-a100.hpcaas>
This commit is contained in:
Martin Gleize
2025-01-19 19:40:40 +01:00
committed by GitHub
parent 81763c58a0
commit bbe5f9de7d
7 changed files with 197 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
tests/weight_loading/models.txt
View File

@ -30,4 +30,5 @@ marlin, nm-testing/zephyr-beta-7b-marlin-g128, main
marlin, robertgshaw2/zephyr-7b-beta-channelwise-marlin, main
qqq, HandH1998/QQQ-Llama-3-8b-g128, main
qqq, HandH1998/QQQ-Llama-3-8b, main
hqq, nm-testing/Llama-3.2-1B-Instruct-HQQ, main
hqq, nm-testing/Llama-3.2-1B-Instruct-HQQ, main
None, mgleize/fairseq2-dummy-Llama-3.2-1B, main

13
tests/weight_loading/test_weight_loading.py
View File

@ -20,12 +20,13 @@ def test_weight_loading(vllm_runner):
"""
Test parameter weight loading with tp>1.
"""
with vllm_runner(model_name=MODEL_NAME,
revision=REVISION,
dtype=torch.half if QUANTIZATION == "gptq" else "auto",
quantization=QUANTIZATION,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=2) as model:
with vllm_runner(
model_name=MODEL_NAME,
revision=REVISION,
dtype=torch.half if QUANTIZATION == "gptq" else "auto",
quantization=None if QUANTIZATION == "None" else QUANTIZATION,
max_model_len=MAX_MODEL_LEN,
tensor_parallel_size=2) as model:
output = model.generate_greedy("Hello world!", max_tokens=20)
print(output)