Update amd docker and nightly models. (#6687)

scripts/amd_ci_start_container.sh

@@ -9,7 +9,7 @@ else
 fi
 
 # Pull the image
-IMAGE="ghcr.io/saienduri/sglang-aiter-backend-v0.1.2:518"
+IMAGE="lmsysorg/sglang:v0.4.6.post5-rocm630"
 echo "Pulling Docker image: $IMAGE"
 docker pull "$IMAGE"
 

test/srt/test_nightly_gsm8k_eval_amd.py

@@ -26,18 +26,17 @@ MODEL_SCORE_THRESHOLDS = {
     "meta-llama/Llama-3.1-70B-Instruct": 0.95,
     "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.64,
     "Qwen/Qwen2-57B-A14B-Instruct": 0.86,
-    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.82,
+    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.83,
     "neuralmagic/Mistral-7B-Instruct-v0.3-FP8": 0.54,
     "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8": 0.94,
     "neuralmagic/Qwen2-72B-Instruct-FP8": 0.94,
     "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.86,
-    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8": 0.61,
+    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8": 0.65,
     "google/gemma-2-27b-it": 0.91,
+    "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8": 0.84,
 }
 
-# Models currently failing on AMD MI300x.
 failing_models = {
-    "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8",
     "neuralmagic/gemma-2-2b-it-FP8",
 }
 
@@ -61,6 +60,16 @@ DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2 = remove_failing_models(
     DEFAULT_MODEL_NAME_FOR_NIGHTLY_EVAL_FP8_TP2
 )
 
+NO_MOE_PADDING_MODELS = {"neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8"}
+DISABLE_HF_XET_MODELS = {
+    "Qwen/Qwen2-57B-A14B-Instruct",
+    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8",
+}
+TRITON_MOE_MODELS = {
+    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8",
+    "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8",
+}
+
 
 def parse_models(model_string):
     return [model.strip() for model in model_string.split(",") if model.strip()]
@@ -156,6 +165,16 @@ class TestNightlyGsm8KEval(unittest.TestCase):
         for model_group, is_fp8, is_tp2 in self.model_groups:
             for model in model_group:
                 with self.subTest(model=model):
+                    os.environ["SGLANG_MOE_PADDING"] = (
+                        "0" if model in NO_MOE_PADDING_MODELS else "1"
+                    )
+                    os.environ["HF_HUB_DISABLE_XET"] = (
+                        "1" if model in DISABLE_HF_XET_MODELS else "0"
+                    )
+                    os.environ["SGLANG_AITER_MOE"] = (
+                        "0" if model in TRITON_MOE_MODELS else "1"
+                    )
+
                     process = popen_launch_server_wrapper(self.base_url, model, is_tp2)
 
                     args = SimpleNamespace(