[AMD] rocm 7.2 image release, PR test, Nightly Test (#17799)

Co-authored-by: Alan Kao <akao@amd.com> Co-authored-by: bingxche <Bingxu.Chen@amd.com> Co-authored-by: Michael <13900043+michaelzhang-ai@users.noreply.github.com>
2026-02-11 21:29:25 -08:00
parent 93ede0db19
commit 20554a0a4f
25 changed files with 2716 additions and 153 deletions
@@ -68,7 +68,7 @@ GPT_OSS_MODELS = [
            "triton",
            "--trust-remote-code",
        ],
-        env_vars={"SGLANG_USE_AITER": "0"},
+        env_vars={"SGLANG_USE_AITER": "1"},
    ),
    ModelConfig(
        model_path="lmsys/gpt-oss-120b-bf16",
@@ -86,7 +86,7 @@ GPT_OSS_MODELS = [
            "triton",
            "--trust-remote-code",
        ],
-        env_vars={"SGLANG_USE_AITER": "0"},
+        env_vars={"SGLANG_USE_AITER": "1"},
    ),
 ]

@@ -41,7 +41,7 @@ MODEL_SCORE_THRESHOLDS = {
    # Llama 3.2 series (smaller models)
    "meta-llama/Llama-3.2-3B-Instruct": 0.55,
    # Mistral series
-    "mistralai/Mistral-7B-Instruct-v0.3": 0.58,
+    "mistralai/Mistral-7B-Instruct-v0.3": 0.55,
    "mistralai/Mixtral-8x7B-Instruct-v0.1": 0.61,
    # DeepSeek series
    "deepseek-ai/DeepSeek-Coder-V2-Lite-Instruct": 0.85,
@@ -108,10 +108,10 @@ DISABLE_HF_XET_MODELS = {
    "neuralmagic/Qwen2-57B-A14B-Instruct-FP8",
 }
 TRITON_MOE_MODELS = {
-    "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8",
+    # "neuralmagic/Mixtral-8x7B-Instruct-v0.1-FP8",
    "neuralmagic/DeepSeek-Coder-V2-Lite-Instruct-FP8",
-    "mistralai/Mixtral-8x7B-Instruct-v0.1",
-    "mistralai/Mistral-7B-Instruct-v0.3",
+    # "mistralai/Mixtral-8x7B-Instruct-v0.1",
+    # "mistralai/Mistral-7B-Instruct-v0.3",
 }
 # AMD-specific models that need special launch config (matching in-house CI sanity_check.py)
 # AMD_SPECIAL_CONFIG_MODELS = {
@@ -120,9 +120,9 @@ AMD_VLM_MODELS = [

 # Models that need special handling on AMD (MoE models)
 TRITON_ATTENTION_MODELS = {
-    "deepseek-ai/deepseek-vl2-small",
-    "Qwen/Qwen3-VL-30B-A3B-Instruct",
-    "moonshotai/Kimi-VL-A3B-Instruct",
+    # "deepseek-ai/deepseek-vl2-small",
+    # "Qwen/Qwen3-VL-30B-A3B-Instruct",
+    # "moonshotai/Kimi-VL-A3B-Instruct",
 }

 # Models known to fail on AMD - exclude from testing
@@ -32,9 +32,9 @@ from sglang.test.test_utils import (
 )
 from sglang.utils import download_and_cache_file, read_jsonl

-# Register for AMD CI - MI35x DeepSeek-V3.2 accuracy test (~60 min for basic only)
+# Register for AMD CI - MI35x DeepSeek-V3.2 accuracy test (~90 min for basic only)
 register_amd_ci(
-    est_time=3600,
+    est_time=5400,
    suite="nightly-amd-8-gpu-mi35x-deepseek-v32",
    nightly=True,
 )
@@ -74,7 +74,7 @@ MI35X_DEEPSEEK_V32_MODELS = [
        model_path="deepseek-ai/DeepSeek-V3.2",
        tp_size=8,
        accuracy_threshold=0.93,
-        timeout=3600,
+        timeout=5400,
        variant="basic",
        other_args=[
            "--trust-remote-code",
@@ -22,7 +22,6 @@ from sglang.test.ci.ci_register import register_amd_ci
 from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
 from sglang.test.send_one import BenchArgs, send_one_prompt
 from sglang.test.test_utils import (
-    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
    DEFAULT_URL_FOR_TEST,
    CustomTestCase,
    is_in_ci,
@@ -32,7 +31,7 @@ from sglang.test.test_utils import (

 # Register for AMD CI - MI35x DeepSeek-V3.2 TP+MTP accuracy test
 register_amd_ci(
-    est_time=3600,
+    est_time=5400,
    suite="nightly-amd-accuracy-8-gpu-mi35x-deepseek-v32-mtp",
    nightly=True,
 )
@@ -55,10 +54,15 @@ class TestDeepseekV32TPMTP(CustomTestCase):
    def setUpClass(cls):
        cls.model = DEEPSEEK_V32_MODEL_PATH
        cls.base_url = DEFAULT_URL_FOR_TEST
+        # Use same args as perf test (which passes successfully)
        other_args = [
            "--trust-remote-code",
            "--tp",
            "8",
+            "--nsa-prefill-backend",
+            "tilelang",
+            "--nsa-decode-backend",
+            "tilelang",
            "--speculative-algorithm",
            "EAGLE",
            "--speculative-num-steps",
@@ -67,19 +71,17 @@ class TestDeepseekV32TPMTP(CustomTestCase):
            "1",
            "--speculative-num-draft-tokens",
            "4",
-            "--mem-frac",
+            "--mem-fraction-static",
            "0.7",
            "--model-loader-extra-config",
            '{"enable_multithread_load": true}',
-            "--nsa-prefill-backend",
-            "tilelang",
-            "--nsa-decode-backend",
-            "tilelang",
+            "--watchdog-timeout",
+            "1200",
        ]
        cls.process = popen_launch_server(
            cls.model,
            cls.base_url,
-            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            timeout=5400,
            other_args=other_args,
        )

@@ -97,8 +99,8 @@ class TestDeepseekV32TPMTP(CustomTestCase):
        args = SimpleNamespace(
            num_shots=20,
            data_path=None,
-            num_questions=1400,
-            parallel=1400,
+            num_questions=200,
+            parallel=64,
            max_new_tokens=512,
            host="http://127.0.0.1",
            port=int(self.base_url.split(":")[-1]),
@@ -75,9 +75,7 @@ MI35X_GPT_OSS_MODELS = [
            "triton",
            "--trust-remote-code",
        ],
-        env_vars={
-            "SGLANG_USE_AITER": "0"
-        },  # Disabled due to SWA eviction bug with aiter (#17220)
+        env_vars={"SGLANG_USE_AITER": "1"},
    ),
    ModelConfig(
        model_path="openai/gpt-oss-120b",
@@ -95,9 +93,7 @@ MI35X_GPT_OSS_MODELS = [
            "triton",
            "--trust-remote-code",
        ],
-        env_vars={
-            "SGLANG_USE_AITER": "0"
-        },  # Disabled due to SWA eviction bug with aiter (#17220)
+        env_vars={"SGLANG_USE_AITER": "1"},
    ),
 ]

@@ -115,6 +115,7 @@ class TestNightlyDeepseekV32BasicPerformance(unittest.TestCase):
                variant=self.variant_config["name"],
                extra_bench_args=["--trust-remote-code"],
                enable_profile=False,  # Disable profiling for AMD tests
+                timeout=5400,  # Extended timeout for large model loading
            )
            results = result_tuple[0]
            success = result_tuple[1]