Add weight loading test for fairseq2

Signed-off-by: Martin Gleize <[email protected]>

Author: MartinGleize

tests/weight_loading/models.txt

@@ -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

tests/weight_loading/test_weight_loading.py

@@ -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)