Add Qwen 3.8 27B Megatron support

megatron_runtime/pyproject.toml

@@ -22,7 +22,7 @@ dependencies = [
     "nvidia-ml-py==13.580.82",
     "nvidia-modelopt>=0.42.0a0",
     "nvidia-resiliency-ext<0.5",
-    "openai>=2.14.0",
+    "openai>=2.14.0,<3",
     "peft>=0.14.0",
     "polars>=1.26.0",
     "pydantic>=2.12",

pyproject.toml

@@ -7,7 +7,7 @@ requires-python = ">=3.12"
 dependencies = [
     "aiohttp>=3.10.0",
     "anthropic>=0.77.0",
-    "openai>=2.14.0",
+    "openai>=2.14.0,<3",
     "pydantic>=2.12",
     "requests>=2.32.0",
     "typing-extensions>=4.13",

src/art/megatron/model_support/registry.py

@@ -161,6 +161,7 @@
         "Qwen/Qwen3.5-4B",
         "Qwen/Qwen3.5-27B",
         "Qwen/Qwen3.6-27B",
+        "Qwen/Qwen3.8-27B",
     ),
     default_target_modules=_QWEN3_5_DENSE_TARGET_MODULES,
     native_vllm_lora_status=_VALIDATED_NATIVE_VLLM_LORA_STATUS,

src/art/trainer_rank/_impl.py

@@ -1417,7 +1417,7 @@ def _validate_checkpoint_adapter_config(
             raise TypeError(
                 "adapter_config['base_model_name_or_path'] must be a string"
             )
-        if base_model.startswith(("Qwen/Qwen3.5-", "Qwen/Qwen3.6-")):
+        if base_model.startswith(("Qwen/Qwen3.5-", "Qwen/Qwen3.6-", "Qwen/Qwen3.8-")):
             dimensions = {
                 "num_attention_heads": getattr(
                     self.runtime.provider, "num_attention_heads", None

tests/integration/megatron/model_support/gdn_fp32_reference.py

@@ -224,8 +224,10 @@ def install_megatron_qwen35_gdn_fp32_reference(
     *,
     base_model: str,
 ) -> None:
-    model_key = base_model.lower()
-    if "qwen3.5" not in model_key and "qwen3_5" not in model_key:
+    from art.megatron.model_support.registry import get_model_support_handler
+
+    handler = get_model_support_handler(base_model)
+    if handler.key not in {"qwen3_5_dense", "qwen3_5_moe"}:
         return
     from art.megatron.gdn import operator as gdn_operator
 

tests/integration/megatron/model_support/hf_parity_worker.py

@@ -1109,8 +1109,10 @@ def _hf_prefix_tree_paths(
 
 
 def _install_hf_qwen35_gdn_fp32_reference(model: Any, *, base_model: str) -> None:
-    model_key = base_model.lower()
-    if "qwen3.5" not in model_key and "qwen3_5" not in model_key:
+    from art.megatron.model_support.registry import get_model_support_handler
+
+    handler = get_model_support_handler(base_model)
+    if handler.key not in {"qwen3_5_dense", "qwen3_5_moe"}:
         return
     patched = 0
     for module in model.modules():

tests/integration/megatron/model_support/test_provider_support.py

@@ -137,6 +137,16 @@ def test_openpipe_qwen3_14b_instruct_uses_qwen3_dense_support() -> None:
     assert handler.key == "qwen3_dense"
 
 
+def test_qwen38_27b_uses_qwen35_dense_support() -> None:
+    spec = get_model_support_spec("Qwen/Qwen3.8-27B")
+    handler = get_model_support_handler("Qwen/Qwen3.8-27B")
+
+    assert spec.key == "qwen3_5_dense"
+    assert spec.is_moe is False
+    assert spec.native_vllm_lora_status == "validated"
+    assert handler.key == "qwen3_5_dense"
+
+
 def test_meta_llama_32_1b_instruct_uses_llama3_dense_support() -> None:
     model = "meta-llama/Llama-3.2-1B-Instruct"
     spec = get_model_support_spec(model)

tests/integration/megatron/model_support/test_workflow.py

@@ -885,6 +885,22 @@ def test_validated_architecture_representative_models_are_fixed() -> None:
     ]
 
 
+def test_qwen38_uses_its_measured_throughput_fingerprint() -> None:
+    qwen35 = handler_workflow_resources_for_base_model("Qwen/Qwen3.5-27B")
+    qwen38 = handler_workflow_resources_for_base_model("Qwen/Qwen3.8-27B")
+    assert qwen35 is not None and qwen35.e2e_throughput is not None
+    assert qwen38 is not None and qwen38.e2e_throughput is not None
+    qwen35_config = qwen35.e2e_throughput.throughput
+    qwen38_config = qwen38.e2e_throughput.throughput
+    assert qwen35_config is not None and qwen38_config is not None
+    assert qwen35_config.thresholds["b300"].calibration_fingerprint == (
+        "5617e8880591545a3281ff14d1fe5197eeefc21a81ec80d1a107fd31421d37a0"
+    )
+    assert qwen38_config.thresholds["b300"].calibration_fingerprint == (
+        "b07ee7ec6338ec021463a43a90fc96c5c5a036b4a04d90b80e1d22c1eef86774"
+    )
+
+
 def test_dsv4_runtime_stages_use_full_model_resources() -> None:
     resources = handler_workflow_resources_for_base_model(
         "deepseek-ai/DeepSeek-V4-Flash"

tests/integration/megatron/model_support/workflow_fixtures.py

@@ -52,6 +52,7 @@
     "Qwen/Qwen3-32B": "9216db5781bf21249d130ec9da846c4624c16137",
     "Qwen/Qwen3-30B-A3B": "ad44e777bcd18fa416d9da3bd8f70d33ebb85d39",
     "Qwen/Qwen3.5-27B": "fc05daec18b0a78c049392ed2e771dde82bdf654",
+    "Qwen/Qwen3.8-27B": "1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0",
     "Qwen/Qwen3.5-35B-A3B": "59d61f3ce65a6d9863b86d2e96597125219dc754",
     "google/gemma-4-31B-it": "842da3794eaa0b77d5f08bae87a17459d91ff475",
     "google/gemma-4-26B-A4B-it": "4d7ae4984b7db7de8f8457170b3f1a419ee76d52",

tests/integration/megatron/model_support/workflow_resources.py

@@ -425,6 +425,9 @@ class HandlerWorkflowResources(BaseModel):
         (81_700, 76_400, 4_850, 0.88, 2.5),
     ),
 }
+_B300_THROUGHPUT_FINGERPRINT_OVERRIDES = {
+    "Qwen/Qwen3.8-27B": "b07ee7ec6338ec021463a43a90fc96c5c5a036b4a04d90b80e1d22c1eef86774",
+}
 _H200_THROUGHPUT_FLOORS = {
     "llama3_dense": (18_300, 17_200, 4_400, 0.89, 7.0),
     "qwen3_dense": (24_100, 23_100, 5_000, 0.91, 7.0),
@@ -554,7 +557,22 @@ def handler_workflow_resources_for_base_model(
         base_model,
         allow_unvalidated_arch=allow_unvalidated_arch,
     )
-    return HANDLER_WORKFLOW_RESOURCES.get(spec.handler_key)
+    resources = HANDLER_WORKFLOW_RESOURCES.get(spec.handler_key)
+    fingerprint = _B300_THROUGHPUT_FINGERPRINT_OVERRIDES.get(base_model)
+    if resources is None or resources.e2e_throughput is None or fingerprint is None:
+        return resources
+    stage = resources.e2e_throughput
+    config = stage.throughput
+    if config is None:
+        raise RuntimeError(f"missing throughput config for {base_model}")
+    thresholds = dict(config.thresholds)
+    thresholds["b300"] = thresholds["b300"].model_copy(
+        update={"calibration_fingerprint": fingerprint}
+    )
+    config = config.model_copy(update={"thresholds": thresholds})
+    return resources.model_copy(
+        update={"e2e_throughput": stage.model_copy(update={"throughput": config})}
+    )
 
 
 def _h200_equivalent_slots_for_total_gib(total_gib: float) -> int:

tests/unit/test_trainer_rank_validation.py

@@ -978,7 +978,13 @@ def test_checkpoint_slot_adapter_config_is_validated_and_copied() -> None:
         trainer._validate_checkpoint_adapter_config("student", {"r": 8}, alpha=None)
 
 
-def test_qwen35_checkpoint_adapter_config_captures_attention_dimensions() -> None:
+@pytest.mark.parametrize(
+    "base_model",
+    ("Qwen/Qwen3.5-4B", "Qwen/Qwen3.6-27B", "Qwen/Qwen3.8-27B"),
+)
+def test_qwen35_checkpoint_adapter_config_captures_attention_dimensions(
+    base_model: str,
+) -> None:
     runtime = _runtime()
     runtime.provider.num_attention_heads = 16
     runtime.provider.num_query_groups = 4
@@ -988,7 +994,7 @@ def test_qwen35_checkpoint_adapter_config_captures_attention_dimensions() -> Non
     retained = trainer._validate_checkpoint_adapter_config(
         "student",
         {
-            "base_model_name_or_path": "Qwen/Qwen3.5-4B",
+            "base_model_name_or_path": base_model,
             "r": 8,
             "lora_alpha": 16,
             "target_modules": ["q_proj"],