Update nightly thresholds and dependencies. (#6635)

scripts/amd_ci_install_dependency.sh

@@ -13,3 +13,5 @@ docker exec -w /human-eval ci_sglang pip install -e .
 docker exec -w / ci_sglang mkdir -p /dummy-grok
 mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
 docker cp ./dummy-grok ci_sglang:/
+
+docker exec ci_sglang pip install huggingface_hub[hf_xet]

test/srt/test_nightly_gsm8k_eval_amd.py

@@ -21,24 +21,24 @@ from sglang.test.test_utils import (
 
 MODEL_SCORE_THRESHOLDS = {
     "meta-llama/Llama-3.1-8B-Instruct": 0.82,
-    "mistralai/Mistral-7B-Instruct-v0.3": 0.56,
+    "mistralai/Mistral-7B-Instruct-v0.3": 0.58,
     "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": 0.85,
     "meta-llama/Llama-3.1-70B-Instruct": 0.95,
-    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.62,
+    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.64,
     "Qwen/Qwen2-57B-A14B-Instruct": 0.86,
-    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.81,
+    "neuralmagic/Meta-Llama-3.1-8B-Instruct-FP8": 0.82,
     "neuralmagic/Mistral-7B-Instruct-v0.3-FP8": 0.54,
-    "neuralmagic/Meta-Llama-3.1-70B-Instruct-FP8": 0.93,
-    "neuralmagic/Qwen2-72B-Instruct-FP8": 0.93,
-    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8": 0.82,
+    "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,
+    "google/gemma-2-27b-it": 0.91,
 }
 
 # Models currently failing on AMD MI300x.
 failing_models = {
-    "google/gemma-2-27b-it",
     "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8",
     "neuralmagic/gemma-2-2b-it-FP8",
-    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8",
 }