Add Mistral Large 3 to nightly CI tests (#14459)

.github/workflows/nightly-test-nvidia.yml

@@ -420,7 +420,28 @@ jobs:
         run: |
           IS_BLACKWELL=1 bash scripts/ci/ci_install_dependency.sh
 
+      - name: Run Mistral-Large-3 nightly performance test
+        timeout-minutes: 180
+        env:
+          TRACE_BASE_URL: https://raw.githubusercontent.com/sglang-bot/sglang-ci-data/main/traces/${{ github.run_id }}
+          PERFETTO_RELAY_URL: ${{ vars.PERFETTO_RELAY_URL }}
+          GPU_CONFIG: "8-gpu-b200"
+          SGLANG_ENABLE_JIT_DEEPGEMM: "0"
+        run: |
+          rm -rf test/performance_profiles_mistral_large3/
+          cd test
+          IS_BLACKWELL=1 python3 nightly/test_mistral_large3_perf.py
+
+      - name: Publish Mistral-Large-3 traces to storage repo
+        env:
+          GITHUB_TOKEN: ${{ secrets.GH_PAT_FOR_NIGHTLY_CI_DATA }}
+          GITHUB_RUN_ID: ${{ github.run_id }}
+          GITHUB_RUN_NUMBER: ${{ github.run_number }}
+        run: |
+          python3 scripts/ci/publish_traces.py --traces-dir test/performance_profiles_mistral_large3
+
       - name: Run DeepSeek v3.1 nightly performance test
+        if: always()
         timeout-minutes: 180
         env:
           TRACE_BASE_URL: https://raw.githubusercontent.com/sglang-bot/sglang-ci-data/main/traces/${{ github.run_id }}

python/sglang/srt/model_loader/weight_utils.py

@@ -79,13 +79,15 @@ class DisabledTqdm(tqdm):
         super().__init__(*args, **kwargs)
 
 
-def get_lock(model_name_or_path: str, cache_dir: Optional[str] = None):
+def get_lock(
+    model_name_or_path: str, cache_dir: Optional[str] = None, suffix: str = ""
+):
     lock_dir = cache_dir or temp_dir
     os.makedirs(os.path.dirname(lock_dir), exist_ok=True)
     model_name = model_name_or_path.replace("/", "-")
     hash_name = hashlib.sha256(model_name.encode()).hexdigest()
     # add hash to avoid conflict with old users' lock files
-    lock_file_name = hash_name + model_name + ".lock"
+    lock_file_name = hash_name + model_name + suffix + ".lock"
     # mode 0o666 is required for the filelock to be shared across users
     lock = filelock.FileLock(os.path.join(lock_dir, lock_file_name), mode=0o666)
     return lock
@@ -309,8 +311,22 @@ def find_local_hf_snapshot_dir(
         except Exception as e:
             logger.warning("Failed to find local snapshot in default HF cache: %s", e)
 
-    # Check for incomplete files and clean up if found
+    # if local snapshot exists, validate it contains at least one weight file
-    if found_local_snapshot_dir:
+    # matching allow_patterns before skipping download.
+    if found_local_snapshot_dir is None:
+        return None
+
+    # Use file lock to prevent multiple processes (TP ranks) from
+    # validating and cleaning up the same model cache simultaneously.
+    # This prevents race conditions where multiple ranks detect corruption
+    # and try to delete the same files at the same time.
+    with get_lock(model_name_or_path, cache_dir, suffix="-validation"):
+        # Re-check if snapshot dir still exists after acquiring lock
+        # (another process may have already cleaned it up)
+        if not os.path.isdir(found_local_snapshot_dir):
+            return None
+
+        # Check for incomplete files and clean up if found
         repo_folder = os.path.abspath(
             os.path.join(found_local_snapshot_dir, "..", "..")
         )
@@ -334,91 +350,90 @@ def find_local_hf_snapshot_dir(
             )
             return None
 
-    # if local snapshot exists, validate it contains at least one weight file
+        local_weight_files: List[str] = []
-    # matching allow_patterns before skipping download.
+        try:
-    if found_local_snapshot_dir is None:
+            for pattern in allow_patterns:
-        return None
+                matched_files = glob.glob(
-
+                    os.path.join(found_local_snapshot_dir, pattern)
-    local_weight_files: List[str] = []
-    try:
-        for pattern in allow_patterns:
-            matched_files = glob.glob(os.path.join(found_local_snapshot_dir, pattern))
-            for f in matched_files:
-                # os.path.exists returns False for broken symlinks.
-                if not os.path.exists(f):
-                    continue
-                local_weight_files.append(f)
-    except Exception as e:
-        logger.warning(
-            "Failed to scan local snapshot %s with patterns %s: %s",
-            found_local_snapshot_dir,
-            allow_patterns,
-            e,
-        )
-        local_weight_files = []
-
-    # Validate sharded models and check for corruption
-    if local_weight_files:
-        is_valid, error_msg, corrupted_files = _validate_sharded_model(
-            found_local_snapshot_dir, local_weight_files
-        )
-        if not is_valid:
-            if corrupted_files:
-                # Selective cleanup: only remove corrupted files
-                log_info_on_rank0(
-                    logger,
-                    f"Found {len(corrupted_files)} corrupted file(s) for "
-                    f"{model_name_or_path}: {error_msg}. "
-                    "Will selectively clean and re-download only these files.",
                 )
-                _cleanup_corrupted_files_selective(model_name_or_path, corrupted_files)
+                for f in matched_files:
-                return None
+                    # os.path.exists returns False for broken symlinks.
-            else:
+                    if not os.path.exists(f):
-                # Cannot selectively clean (e.g., missing shards) - remove entire cache
+                        continue
-                log_info_on_rank0(
+                    local_weight_files.append(f)
-                    logger,
+        except Exception as e:
-                    f"Validation failed for {model_name_or_path}: {error_msg}. "
+            logger.warning(
-                    "Will remove entire cache and re-download.",
+                "Failed to scan local snapshot %s with patterns %s: %s",
-                )
+                found_local_snapshot_dir,
-                _cleanup_corrupted_model_cache(
+                allow_patterns,
-                    model_name_or_path, found_local_snapshot_dir, error_msg
+                e,
-                )
+            )
-                return None
+            local_weight_files = []
 
-        # Also validate single (non-sharded) safetensors files
+        # Validate sharded models and check for corruption
-        for f in local_weight_files:
+        if local_weight_files:
-            base_name = os.path.basename(f)
+            is_valid, error_msg, corrupted_files = _validate_sharded_model(
-            # Check if this is a single model file (not sharded)
+                found_local_snapshot_dir, local_weight_files
-            # Include adapter_model.safetensors for LoRA adapters
+            )
-            if base_name in [
+            if not is_valid:
-                "model.safetensors",
+                if corrupted_files:
-                "pytorch_model.safetensors",
+                    # Selective cleanup: only remove corrupted files
-                "adapter_model.safetensors",
-            ]:
-                if not _validate_safetensors_file(f):
                     log_info_on_rank0(
                         logger,
-                        f"Corrupted model file {base_name} for {model_name_or_path}. "
+                        f"Found {len(corrupted_files)} corrupted file(s) for "
-                        "Will selectively clean and re-download this file.",
+                        f"{model_name_or_path}: {error_msg}. "
+                        "Will selectively clean and re-download only these files.",
+                    )
+                    _cleanup_corrupted_files_selective(
+                        model_name_or_path, corrupted_files
+                    )
+                    return None
+                else:
+                    # Cannot selectively clean (e.g., missing shards) - remove entire cache
+                    log_info_on_rank0(
+                        logger,
+                        f"Validation failed for {model_name_or_path}: {error_msg}. "
+                        "Will remove entire cache and re-download.",
+                    )
+                    _cleanup_corrupted_model_cache(
+                        model_name_or_path, found_local_snapshot_dir, error_msg
                     )
-                    # Selective cleanup for single file
-                    _cleanup_corrupted_files_selective(model_name_or_path, [f])
                     return None
 
-    if len(local_weight_files) > 0:
+            # Also validate single (non-sharded) safetensors files
-        log_info_on_rank0(
+            for f in local_weight_files:
-            logger,
+                base_name = os.path.basename(f)
-            f"Found local HF snapshot for {model_name_or_path} at "
+                # Check if this is a single model file (not sharded)
-            f"{found_local_snapshot_dir}; skipping download.",
+                # Include adapter_model.safetensors for LoRA adapters
-        )
+                if base_name in [
-        return found_local_snapshot_dir
+                    "model.safetensors",
-    else:
+                    "pytorch_model.safetensors",
-        log_info_on_rank0(
+                    "adapter_model.safetensors",
-            logger,
+                ]:
-            f"Local HF snapshot at {found_local_snapshot_dir} has no files matching "
+                    if not _validate_safetensors_file(f):
-            f"{allow_patterns}; will attempt download.",
+                        log_info_on_rank0(
-        )
+                            logger,
-    return None
+                            f"Corrupted model file {base_name} for {model_name_or_path}. "
+                            "Will selectively clean and re-download this file.",
+                        )
+                        # Selective cleanup for single file
+                        _cleanup_corrupted_files_selective(model_name_or_path, [f])
+                        return None
+
+        if len(local_weight_files) > 0:
+            log_info_on_rank0(
+                logger,
+                f"Found local HF snapshot for {model_name_or_path} at "
+                f"{found_local_snapshot_dir}; skipping download.",
+            )
+            return found_local_snapshot_dir
+        else:
+            log_info_on_rank0(
+                logger,
+                f"Local HF snapshot at {found_local_snapshot_dir} has no files matching "
+                f"{allow_patterns}; will attempt download.",
+            )
+            return None
 
 
 def download_weights_from_hf(

test/nightly/test_mistral_large3_perf.py

@@ -0,0 +1,105 @@
+import os
+import unittest
+from types import SimpleNamespace
+
+from nightly_utils import NightlyBenchmarkRunner
+
+from sglang.srt.utils import kill_process_tree
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.run_eval import run_eval
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    _parse_int_list_env,
+    popen_launch_server,
+)
+
+register_cuda_ci(est_time=600, suite="nightly-8-gpu-b200", nightly=True)
+
+MISTRAL_LARGE3_MODEL_PATH = "mistralai/Mistral-Large-3-675B-Instruct-2512"
+PROFILE_DIR = "performance_profiles_mistral_large3"
+
+
+class TestNightlyMistralLarge3Performance(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        # Set environment variable to disable JIT DeepGemm
+        os.environ["SGLANG_ENABLE_JIT_DEEPGEMM"] = "0"
+
+        cls.model = MISTRAL_LARGE3_MODEL_PATH
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.batch_sizes = [1, 1, 8, 16, 64]
+        cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
+        cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
+
+        # Mistral-Large-3-675B requires TP=8 and trtllm_mla attention backend
+        cls.other_args = [
+            "--tp",
+            "8",
+            "--attention-backend",
+            "trtllm_mla",
+            "--model-loader-extra-config",
+            '{"enable_multithread_load": true}',
+            "--chat-template",
+            "mistral",
+        ]
+
+        cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
+        cls.runner.setup_profile_directory()
+
+    @classmethod
+    def tearDownClass(cls):
+        # Clean up environment variable
+        if "SGLANG_ENABLE_JIT_DEEPGEMM" in os.environ:
+            del os.environ["SGLANG_ENABLE_JIT_DEEPGEMM"]
+
+    def test_bench_one_batch(self):
+        results, success = self.runner.run_benchmark_for_model(
+            model_path=self.model,
+            batch_sizes=self.batch_sizes,
+            input_lens=self.input_lens,
+            output_lens=self.output_lens,
+            other_args=self.other_args,
+        )
+
+        self.runner.add_report(results)
+        self.runner.write_final_report()
+
+        if not success:
+            raise AssertionError(
+                f"Benchmark failed for {self.model}. Check the logs for details."
+            )
+
+    def test_accuracy_mgsm(self):
+        """Run MGSM accuracy evaluation for Mistral Large 3."""
+        process = popen_launch_server(
+            model=self.model,
+            base_url=self.base_url,
+            other_args=self.other_args,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+        )
+
+        try:
+            args = SimpleNamespace(
+                base_url=self.base_url,
+                model=self.model,
+                eval_name="mgsm_en",
+                num_examples=None,
+                num_threads=1024,
+            )
+            metrics = run_eval(args)
+            print(f"MGSM accuracy for {self.model}: {metrics['score']}")
+
+            # Placeholder threshold - adjust after first successful run
+            expected_threshold = 0.90
+            self.assertGreaterEqual(
+                metrics["score"],
+                expected_threshold,
+                f"MGSM accuracy {metrics['score']} below threshold {expected_threshold}",
+            )
+        finally:
+            kill_process_tree(process.pid)
+
+
+if __name__ == "__main__":
+    unittest.main()