Merge branch 'main' into fix/mamba-gate-shift-silu-into-ssm

Author: wdykas

examples/post_training/modelopt/convert_model.py

@@ -27,6 +27,7 @@
 from megatron.training import get_args
 from megatron.training.checkpointing import save_checkpoint
 from megatron.training.initialize import initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from megatron.training.utils import print_rank_0, unwrap_model
 from model_provider import model_provider
 
@@ -102,14 +103,12 @@ def check_arguments():
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_convert_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_convert_args, args_defaults={
             'tokenizer_type': 'HuggingFaceTokenizer',
             'no_load_rng': True,
             'no_load_optim': True,
-        },
-    )
+        })
+    initialize_megatron()
     check_arguments()
 
     args = get_args()

examples/post_training/modelopt/export.py

@@ -18,6 +18,7 @@
 from megatron.post_training.model_builder import modelopt_gpt_hybrid_builder
 from megatron.training import get_args, get_model
 from megatron.training.initialize import initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from megatron.training.utils import unwrap_model
 from model_provider import model_provider
 
@@ -49,14 +50,12 @@ def add_modelopt_export_args(parser):
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_modelopt_export_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_modelopt_export_args, args_defaults={
             'tokenizer_type': 'HuggingFaceTokenizer',
             'no_load_rng': True,
             'no_load_optim': True,
-        },
-    )
+        })
+    initialize_megatron()
 
     args = get_args()
 

examples/post_training/modelopt/finetune.py

@@ -484,12 +484,16 @@ def forward_step(data_iterator, model: GPTModel):
 
 
 if __name__ == "__main__":
+    from megatron.training.arguments import parse_and_validate_args
+
+    parse_and_validate_args(
+        extra_args_provider=add_finetune_args,
+        args_defaults={"tokenizer_type": "HuggingFaceTokenizer"},
+    )
     pretrain(
         train_valid_test_sft_datasets_provider,
         partial(model_provider, modelopt_gpt_hybrid_builder),
         ModelType.encoder_or_decoder,
         forward_step,
-        extra_args_provider=add_finetune_args,
-        args_defaults={"tokenizer_type": "HuggingFaceTokenizer"},
         non_loss_data_func=non_loss_data_func,
     )

examples/post_training/modelopt/generate.py

@@ -17,6 +17,7 @@
 from megatron.post_training.model_builder import modelopt_gpt_hybrid_builder
 from megatron.post_training.utils import report_current_memory_info, to_empty_if_meta
 from megatron.training import get_args, get_model, initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from utils import get_hf_tokenizer
 from megatron.training.utils import print_rank_0, unwrap_model
 from model_provider import model_provider
@@ -73,14 +74,12 @@ def get_conversations(example):
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_generate_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_generate_args, args_defaults={
             'tokenizer_type': 'HuggingFaceTokenizer',
             'no_load_rng': True,
             'no_load_optim': True,
-        },
-    )
+        })
+    initialize_megatron()
 
     check_arguments()
 

examples/post_training/modelopt/mmlu.py

@@ -20,6 +20,7 @@
 from megatron.post_training.model_builder import modelopt_gpt_hybrid_builder
 from megatron.post_training.utils import report_current_memory_info
 from megatron.training import get_args, get_model, initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from utils import get_hf_tokenizer
 from megatron.training.utils import print_rank_0, unwrap_model
 import modelopt.torch.quantization as mtq
@@ -133,14 +134,12 @@ def generate_prompt(test_example, dev_examples, few_shots=0, no_subject_prompt=F
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_mmlu_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_mmlu_args, args_defaults={
             'tokenizer_type': 'HuggingFaceTokenizer',
             'no_load_rng': True,
             'no_load_optim': True,
-        },
-    )
+        })
+    initialize_megatron()
 
     args = get_args()
     cache = Cache(args.cache_dir)

examples/post_training/modelopt/offline_feature_extract.py

@@ -42,14 +42,12 @@ def extract_feature(dataset, model, output_dir, idx_start, idx_end):
             torch.distributed.barrier()
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_extract_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_extract_args, args_defaults={
             'tokenizer_type': 'HuggingFaceTokenizer',
             'no_load_rng': True,
             'no_load_optim': True,
-        },
-    )
+        })
+    initialize_megatron()
 
     args = get_args()
     tokenizer = get_tokenizer()

examples/post_training/modelopt/prune.py

@@ -33,6 +33,7 @@
     report_current_memory_info,
 )
 from megatron.training import get_args, get_model, initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from utils import get_hf_tokenizer
 from megatron.training.checkpointing import save_checkpoint
 from megatron.training.utils import print_rank_0, unwrap_model
@@ -149,14 +150,12 @@ def get_params(model):
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_prune_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_prune_args, args_defaults={
             "tokenizer_type": "HuggingFaceTokenizer",
             "no_load_rng": True,
             "no_load_optim": True,
-        },
-    )
+        })
+    initialize_megatron()
 
     args = get_args()
     check_arguments(args)

examples/post_training/modelopt/quantize.py

@@ -45,6 +45,7 @@
     report_current_memory_info,
 )
 from megatron.training import get_args, get_model, initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from utils import get_hf_tokenizer
 from megatron.training.checkpointing import save_checkpoint
 from megatron.training.utils import print_rank_0, unwrap_model
@@ -188,13 +189,10 @@ def get_first_layers_disabled_config(config, num_layers: int = 1, num_layers_to_
     """
     config = copy.deepcopy(config)
     quant_cfg = config.get("quant_cfg", {})
-    quant_cfg.update(
-        {
-            functools.partial(
-                _is_first_layers, num_layers=num_layers, num_layers_to_disable=num_layers_to_disable
-            ): {"enable": False}
-        }
+    predicate = functools.partial(
+        _is_first_layers, num_layers=num_layers, num_layers_to_disable=num_layers_to_disable
     )
+    quant_cfg.append({"quantizer_name": predicate, "enable": False})
     config["quant_cfg"] = quant_cfg
     return config
 
@@ -206,13 +204,10 @@ def get_last_layers_disabled_config(config, num_layers: int = 1, num_layers_to_d
     """
     config = copy.deepcopy(config)
     quant_cfg = config.get("quant_cfg", {})
-    quant_cfg.update(
-        {
-            functools.partial(
-                _is_last_layers, num_layers=num_layers, num_layers_to_disable=num_layers_to_disable
-            ): {"enable": False}
-        }
+    predicate = functools.partial(
+        _is_last_layers, num_layers=num_layers, num_layers_to_disable=num_layers_to_disable
     )
+    quant_cfg.append({"quantizer_name": predicate, "enable": False})
     config["quant_cfg"] = quant_cfg
     return config
 
@@ -224,28 +219,34 @@ def get_modelopt_torch_quantization_config():
         raise ValueError(f"Unsupported quantization config {args.export_quant_cfg}.")
     mtq_config = QUANT_CFG_CHOICES[args.export_quant_cfg]
 
-    fp8_config = {"enable": True, "num_bits": (4, 3), "axis": None}
+    if isinstance(mtq_config["quant_cfg"], dict):
+        # Normalize old dict format to new list format
+        mtq_config["quant_cfg"] = mtq.normalize_quant_cfg_list(mtq_config["quant_cfg"])
+
+    fp8_config = {"enable": True, "cfg": {"num_bits": (4, 3), "axis": None}}
     fp4_config = {
-        "num_bits": (2, 1),
-        "block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
-        "axis": None,
         "enable": True,
+        "cfg": {
+            "num_bits": (2, 1),
+            "block_sizes": {-1: 16, "type": "dynamic", "scale_bits": (4, 3)},
+            "axis": None,
+        },
     }
     if args.export_quant_cfg == "FP8_DEFAULT_CFG":
         # Enable Medusa heads and kv-cache quantization
-        mtq_config["quant_cfg"]["*medusa_heads**"] = fp8_config
+        mtq_config["quant_cfg"].append({"quantizer_name": "*medusa_heads**", **fp8_config})
     if "FP4" in args.export_quant_cfg:
         # Enable Medusa heads and kv-cache quantization
-        mtq_config["quant_cfg"]["*medusa_heads**"] = fp4_config
+        mtq_config["quant_cfg"].append({"quantizer_name": "*medusa_heads**", **fp4_config})
     if "AWQ" in args.export_quant_cfg:
-        weight_quantizer = mtq_config["quant_cfg"]["*weight_quantizer"]  # type: ignore
-        if isinstance(weight_quantizer, list):
-            weight_quantizer = weight_quantizer[0]
-        weight_quantizer["block_sizes"][-1] = 128
-
+        try:
+            weight_quantizer = mtq.find_quant_cfg_entry_by_path(mtq_config["quant_cfg"], "*weight_quantizer")
+            weight_quantizer["block_sizes"][-1] = 128
+        except KeyError:
+            weight_quantizer = None
     # Customization
     if args.disable_qkv_quant:
-        mtq_config["quant_cfg"]["*self_attention*"] = {"enable": False}
+        mtq_config["quant_cfg"].append({"quantizer_name": "*self_attention*", "enable": False})
 
     # KV Cache Quantization
     enable_quant_kv_cache = args.export_kv_cache_quant != "none"
@@ -257,7 +258,7 @@ def get_modelopt_torch_quantization_config():
 
     # Weight Only Quantization
     if args.weight_only:
-        mtq_config["quant_cfg"]["*input_quantizer"] = {"enable": False}
+        mtq_config["quant_cfg"].append({"quantizer_name": "*input_quantizer", "enable": False})
     if args.num_first_layers_to_skip_quant is not None:
         mtq_config = get_first_layers_disabled_config(
             mtq_config,
@@ -346,14 +347,12 @@ def get_calib_dataloader(
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_text_generate_ptq_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_text_generate_ptq_args, args_defaults={
             "tokenizer_type": "HuggingFaceTokenizer",
             "no_load_rng": True,
             "no_load_optim": True,
-        },
-    )
+        })
+    initialize_megatron()
 
     check_arguments()
 

examples/post_training/modelopt/validate.py

@@ -17,6 +17,7 @@
 from megatron.post_training.model_builder import modelopt_gpt_hybrid_builder
 from megatron.post_training.utils import get_mtbench_chat_data
 from megatron.training import get_args, get_model, initialize_megatron
+from megatron.training.arguments import parse_and_validate_args
 from utils import get_hf_tokenizer
 from megatron.training.utils import print_rank_0, unwrap_model
 from model_provider import model_provider
@@ -89,14 +90,12 @@ def report_current_memory_info():
 
 
 if __name__ == "__main__":
-    initialize_megatron(
-        extra_args_provider=add_ar_validation_args,
-        args_defaults={
+    parse_and_validate_args(extra_args_provider=add_ar_validation_args, args_defaults={
             'tokenizer_type': 'HuggingFaceTokenizer',
             'no_load_rng': True,
             'no_load_optim': True,
-        },
-    )
+        })
+    initialize_megatron()
 
     check_arguments()
 

skills/nightly-sync/SKILL.md

@@ -281,12 +281,16 @@ the branch.
      in a collapsed `<details>` block. If git is too old for `--remerge-diff`,
      note the git version and describe the merge strategy used instead.
 - Save the PR number for later phases
-- **Add the `Run functional tests` label** to the PR immediately after
-  creation. This ensures `/ok to test` triggers the full CI suite (unit tests
-  + functional/integration tests with 100-step training and golden value
-  comparison). Without this label, only a lightweight subset runs.
+- **Add the `Run functional tests` and `Run MBridge tests` labels** to the
+  PR immediately after creation. The `Run functional tests` label ensures
+  `/ok to test` triggers the full CI suite (unit tests + functional/
+  integration tests with 100-step training and golden value comparison).
+  The `Run MBridge tests` label triggers the MBridge test suite. Without
+  these labels, only a lightweight subset runs.
   ```bash
-  gh pr edit <PR_NUMBER> --repo $REPO --add-label "Run functional tests"
+  gh pr edit <PR_NUMBER> --repo $REPO \
+    --add-label "Run functional tests" \
+    --add-label "Run MBridge tests"
   ```
 
 ---
@@ -449,8 +453,10 @@ CI progresses. No separate registration poll needed.
   `/ok to test`. Pushing to them directly breaks the CI trigger
   mechanism. Always push to your own sync branch (e.g.
   `main2dev/<DATE>`) instead.
-- **Do NOT forget the `Run functional tests` label.** Without it,
-  the internal GitLab functional tests do not run.
+- **Do NOT forget the `Run functional tests` and `Run MBridge tests`
+  labels.** Without `Run functional tests`, the internal GitLab
+  functional tests do not run; without `Run MBridge tests`, the
+  MBridge test suite does not run.
 
 ### Failure Investigation