add qwen3

vllm-v0.6.2/tests/weight_loading/test_weight_loading.py

@@ -0,0 +1,34 @@
+import os
+
+import torch
+
+MAX_MODEL_LEN = 1024
+MODEL_NAME = os.environ.get("MODEL_NAME",
+                            "robertgshaw2/zephyr-7b-beta-channelwise-gptq")
+REVISION = os.environ.get("REVISION", "main")
+QUANTIZATION = os.environ.get("QUANTIZATION", "gptq_marlin")
+
+'''
+=============================
+Modify by vllm_mlu
+=============================
+@brief(MODEL_NAME,REVERSION,QUANTIZATION): Just test weight loading. 
+''' 
+MODEL_NAME = "facebook/opt-125m"
+REVISION = None
+QUANTIZATION = None
+
+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:
+
+        output = model.generate_greedy("Hello world!", max_tokens=20)
+        print(output)
+        assert output