Support checkpointing Minitron pruning scores / prune without sorting (NVIDIA#361)

Signed-off-by: Keval Morabia <[email protected]>

Author: kevalmorabia97

.gitlab/tests.yml

@@ -27,7 +27,7 @@ unit:
 ##### GPU Tests #####
 .multi-gpu-tests-default:
   extends: .tests-default
-  timeout: 60m
+  timeout: 90m
   image: nvcr.io/nvidia/pytorch:25.06-py3
   variables:
     GIT_DEPTH: 1000 # For correct version for tests/gpu/torch/quantization/plugins/test_megatron.py

CHANGELOG.rst

@@ -17,6 +17,7 @@ Model Optimizer Changelog (Linux)
 - ``high_precision_dtype`` default to fp16 in ONNX quantization, i.e. quantized output model weights are now FP16 by default.
 - Upgrade TensorRT-LLM dependency to 1.1.0rc2.
 - Support Phi-4-multimodal and Qwen2.5-VL quantized HF checkpoint export in ``examples/vlm_ptq``.
+- Support storing and restoring Minitron pruning activations and scores for re-pruning without running the forward loop again.
 - Add Minitron pruning example for Megatron-LM framework. See ``examples/megatron-lm`` for more details.
 
 0.35 (2025-09-04)

docs/source/guides/3_pruning.rst

@@ -4,7 +4,7 @@ Pruning
 
 .. tip::
 
-    Checkout `Llama 3.1 NeMo Minitron Pruning <https://github.com/NVIDIA-NeMo/NeMo/tree/main/tutorials/llm/llama/pruning-distillation>`_ and
+    Checkout `Qwen 3 NeMo Minitron Pruning & Distillation <https://github.com/NVIDIA-NeMo/NeMo/tree/main/tutorials/llm/qwen/pruning-distillation>`_ and
     `ResNet20 on CIFAR-10 Notebook <https://github.com/NVIDIA/TensorRT-Model-Optimizer/blob/main/examples/pruning/cifar_resnet.ipynb>`_
     for an end-to-end example of pruning.
 

examples/llm_distill/README.md

@@ -144,7 +144,7 @@ Loss balancers:
 
 Checkout the stand-alone distillation script in the [NVIDIA NeMo repository](https://docs.nvidia.com/nemo-framework/user-guide/latest/model-optimization/distillation/distillation.html).
 
-You can also look at the tutorial notebooks [here](https://github.com/NVIDIA-NeMo/NeMo/tree/main/tutorials/llm/llama/pruning-distillation) which showcase the usage of Minitron pruning followed by distillation for Llama 3.1 8B step-by-step in NeMo framework.
+You can also look at the NeMo tutorial notebooks [here](https://github.com/NVIDIA-NeMo/NeMo/tree/main/tutorials/llm/qwen/pruning-distillation) which showcase the usage of Minitron pruning followed by distillation for Qwen 3 8B step-by-step in NeMo framework. Hugging Face models can also be converted to NeMo format and used subsequently as shown in the tutorial.
 
 ## Knowledge Distillation (KD) for HuggingFace Models
 

examples/pruning/README.md

@@ -59,23 +59,27 @@ def forward_loop(model):
     evaluate_and_print_results(model, ...)
 
 
-# Specify the pruning constraints
+# Specify the pruning constraints (Check Support Matrix for available pruning dimensions)
 export_config = {
     "hidden_size": 3072,
     "ffn_hidden_size": 9216,
 }
 
 
 # Run the pruning process
-mtp.prune(
+# Save minitron scores at scores_path so we can re-run pruning with different export configs without running the forward loop again
+# NOTE: Skip scores_path on re-running if you want to change the dataset and re-calibrate
+model, pruning_scores = mtp.prune(
     model,
     mode="mcore_minitron",
     constraints={"export_config": export_config},
     dummy_input=None,  # Not used
-    config={"forward_loop": forward_loop},
+    config={"forward_loop": forward_loop, "scores_path": "modelopt_minitron_scores.pth"},
 )
 ```
 
+If your model parameters are already sorted, you can skip the sorting step by setting `"skip_sorting": True` in `config` instead of passing `forward_loop`.
+
 > [!Note]
 > Fine-tuning / distillation is required after pruning to recover the accuracy. Please refer to pruning [fine-tuning](https://nvidia.github.io/TensorRT-Model-Optimizer/guides/3_pruning.html#pruning-fine-tuning) for more details.
 
@@ -91,11 +95,11 @@ mtp.prune(
 
 ## Examples
 
-### Minitron Pruning for Megatron-LM / NeMo Framework LLMs (e.g. Llama 3.1, Nemotron Nano)
+### Minitron Pruning for Megatron-LM / NeMo Framework LLMs (e.g. Qwen 3, Nemotron Nano)
 
 Checkout the Minitron pruning example for the [Megatron-LM Framework](../megatron-lm/README.md#-pruning) and [NeMo Framework](https://docs.nvidia.com/nemo-framework/user-guide/latest/model-optimization/pruning/pruning.html) which showcases the usage of the powerful Minitron pruning algorithm developed by NVIDIA Research for pruning LLMs like Llama 3.1 8B, Qwen 3 8B, Nemotron Nano 12B v2, etc.
 
-You can also look at the NeMo tutorial notebooks [here](https://github.com/NVIDIA-NeMo/NeMo/tree/main/tutorials/llm/llama/pruning-distillation) which showcase the usage of Minitron pruning followed by distillation for Llama 3.1 8B step-by-step in NeMo framework. Hugging Face models can also be converted to NeMo format and used subsequently as shown in the tutorial.
+You can also look at the NeMo tutorial notebooks [here](https://github.com/NVIDIA-NeMo/NeMo/tree/main/tutorials/llm/qwen/pruning-distillation) which showcase the usage of Minitron pruning followed by distillation for Qwen 3 8B step-by-step in NeMo framework. Hugging Face models can also be converted to NeMo format and used subsequently as shown in the tutorial.
 
 Some of the models pruned using Minitron method followed by distillation and post-training are:
 

modelopt/torch/nas/plugins/megatron.py

@@ -649,6 +649,10 @@ def _layer_imp_forward_hook(self, module, args, kwargs, output) -> None:
         if hidden_states.shape[-1] != self.max_hidden_size:
             return
 
+        # use full precision to avoid overflow
+        hidden_states = hidden_states.to(torch.float32)
+        output = output.to(torch.float32)
+
         with torch.no_grad():
             # Lower cosine_similarity means higher importance hence use 1 - cosine_similarity
             score = 1 - F.cosine_similarity(hidden_states, output, dim=2).mean()
@@ -1234,10 +1238,10 @@ def _emb_layernorm_forward_hook(self, module, input, output) -> None:
         output = output.to(torch.float32)  # use full precision to avoid overflow
         activations = output.abs().mean(dim=0)  # [batch_size, hidden_size]
         activations = activations.pow(2).sum(dim=0)  # [hidden_size]
-        if module not in self._activations:
-            self._activations[module] = activations
+        if id(module) not in self._activations:
+            self._activations[id(module)] = activations
         else:
-            self._activations[module] += activations
+            self._activations[id(module)] += activations
 
     def _estimate_hidden_size_importance(self) -> TracedHp.Importance:
         """Return the activation magnitude-based importance of the hidden_size."""
@@ -1284,16 +1288,14 @@ def modify(
                 mamba_head_dim_divisor=mamba_head_dim_divisor,
             )
 
-    def _export_drop_layers(self) -> None:
-        """Drop layers during export if num_layers hparam is set to a smaller value during pruning."""
+    def _get_layer_scores(self) -> dict[int, torch.Tensor]:
+        """Get the layer scores (1-indexed) from the module."""
         num_layers_hp = self.get_hparam("num_layers")
-        if num_layers_hp.active == num_layers_hp.max:  # no depth pruning
-            return
 
         for layer in self.decoder.layers:
             assert layer._scores > 0, "No scores collected for importance estimation."
 
-        # gather layer scores from all TP regions
+        # gather layer scores from all PP ranks
         layer_scores = {}
         for layer in self.decoder.layers:
             layer_scores[layer.layer_number] = layer._scores
@@ -1302,10 +1304,19 @@ def _export_drop_layers(self) -> None:
             all_pp_layer_scores, layer_scores, group=get_pipeline_model_parallel_group()
         )
         layer_scores = {k: v for d in all_pp_layer_scores for k, v in d.items()}  # type: ignore[attr-defined]
-        print_rank_0(f"Layerwise scores for depth pruning: {layer_scores}")
+        print_rank_0(f"Layerwise scores (1-indexed, higher is better): {layer_scores}")
         assert sorted(layer_scores.keys()) == list(range(1, num_layers_hp.max + 1))  # type: ignore[arg-type]
 
+        return layer_scores
+
+    def _export_drop_layers(self) -> None:
+        """Drop layers during export if num_layers hparam is set to a smaller value during pruning."""
+        num_layers_hp = self.get_hparam("num_layers")
+        if num_layers_hp.active == num_layers_hp.max:  # no depth pruning
+            return
+
         # sort layers by scores and drop the lowest ones
+        layer_scores = self._get_layer_scores()
         sorted_layers = sorted(layer_scores.items(), key=lambda x: x[1], reverse=True)
         layers_to_drop = [layer for layer, _ in sorted_layers[num_layers_hp.active :]]  # type: ignore[misc]
         drop_mcore_language_model_layers(self, layers_to_drop=layers_to_drop)
@@ -1337,6 +1348,47 @@ def freeze(self) -> None:
         for layer in self.decoder.layers:
             layer.freeze()
 
+    def get_activations_and_layer_scores(
+        self,
+    ) -> tuple[list[dict[str, torch.Tensor]], dict[int, torch.Tensor]]:
+        """Get the per-rank activations and layer scores from the module."""
+        local_activations = {}
+        for n, m in self.named_modules():
+            if hasattr(m, "_activations"):
+                local_activations[n] = m._activations
+        activations_per_rank = dist.allgather(
+            local_activations, group=get_pipeline_model_parallel_group()
+        )
+        assert len(activations_per_rank) == get_pipeline_model_parallel_world_size()
+
+        layer_scores = self._get_layer_scores()
+
+        return activations_per_rank, layer_scores
+
+    def set_activations_and_layer_scores(
+        self,
+        activations_per_rank: list[dict[str, torch.Tensor]],
+        layer_scores: dict[int, torch.Tensor],
+    ) -> None:
+        """Set the pre-computed layer_scores and per-rank activations instead of running forward.
+
+        Args:
+            layer_scores: Dict from layer_number (1-indexed) to score.
+            activations_per_rank: List of dicts from module name to activations. Should match PP size.
+        """
+        rank = get_pipeline_model_parallel_rank()
+        pp_size = get_pipeline_model_parallel_world_size()
+        assert len(activations_per_rank) == pp_size, (
+            len(activations_per_rank),
+            activations_per_rank,
+            pp_size,
+        )
+        for layer in self.decoder.layers:
+            layer._scores = layer_scores[layer.layer_number]
+        for n, m in self.named_modules():
+            if hasattr(m, "_activations"):
+                m._activations = activations_per_rank[rank][n]
+
 
 def drop_mcore_language_model_layers(model: nn.Module, *, layers_to_drop: list[int]) -> None:
     """Remove given layers (1-indexed) of the model (works with TP and/or PP).

modelopt/torch/opt/searcher.py

@@ -27,6 +27,7 @@
 from collections.abc import Callable
 from contextlib import nullcontext
 from typing import Any, final
+from warnings import warn
 
 import numpy as np
 import pulp
@@ -239,7 +240,11 @@ def load_search_checkpoint(self) -> bool:
         """Load function for search checkpoint returning indicator whether checkpoint was loaded."""
         # check if checkpoint exists
         checkpoint: str | None = self.config["checkpoint"]
-        if checkpoint is None or not os.path.exists(checkpoint):
+        if checkpoint is None:
+            return False
+        if not os.path.exists(checkpoint):
+            if dist.is_master():
+                warn(f"Checkpoint {checkpoint} does not exist! Initializing from scratch.")
             return False
 
         # iterate through state dict and load keys
@@ -250,14 +255,16 @@ def load_search_checkpoint(self) -> bool:
             setattr(self, key, state)
         return True
 
-    def save_search_checkpoint(self) -> None:
+    def save_search_checkpoint(self, verbose=False) -> None:
         """Save function for search checkpoint."""
         # check if save requirements are satisfied
         checkpoint: str | None = self.config["checkpoint"]
         if checkpoint is None or not dist.is_master():
             return
 
         # save state dict
+        if verbose:
+            print(f"Saving searcher state to {checkpoint}...")
         save_dirname, _ = os.path.split(checkpoint)
         if save_dirname:
             os.makedirs(save_dirname, exist_ok=True)

modelopt/torch/prune/plugins/mcore_minitron.py

@@ -24,13 +24,15 @@
 Actual dynamic module implementations are at :mod:`modelopt.torch.nas.plugins.megatron`.
 """
 
+import copy
+
 import torch
 from pydantic import create_model
 
 # isort: off
 # import nas plugin to check if it is enabled else raises an Exception
 from modelopt.torch.nas.plugins.megatron import *  # noqa: F403
-from modelopt.torch.nas.plugins.megatron import HAS_MAMBA
+from modelopt.torch.nas.plugins.megatron import HAS_MAMBA, _DynamicMCoreLanguageModel
 # isort: on
 
 from modelopt.torch.nas.conversion import NASModeRegistry
@@ -60,22 +62,29 @@
 class MCoreMinitronSearcher(BaseSearcher):
     """Searcher for Minitron pruning algorithm."""
 
+    activations_per_rank: list[dict[str, torch.Tensor]]
+    layer_scores: dict[int, torch.Tensor]
+
     @property
     def default_search_config(self) -> SearchConfig:
         """Get the default config for the searcher."""
-        return {**super().default_search_config, "max_iter_data_loader": 1024}
+        return {
+            **super().default_search_config,
+            "max_iter_data_loader": 1024,
+            "skip_sorting": False,
+            "scores_path": None,
+        }
 
     @property
     def default_state_dict(self) -> SearchStateDict:
-        """Return default state dict."""
-        return {}  # Not used
+        """Return default state dict for importance scores and activations from forward loop."""
+        return {"activations_per_rank": [], "layer_scores": {}}
 
     def sanitize_search_config(self, config: SearchConfig | None) -> SearchConfig:
         """Sanitize the search config dict."""
         config = super().sanitize_search_config(config)
-        assert config["data_loader"] or config["forward_loop"], (
-            "Data loader or forward loop must be provided for importance estimation!"
-        )
+        config["checkpoint"] = config["scores_path"]
+        config["verbose"] = True  # Print for all ranks
         return config
 
     def before_search(self) -> None:
@@ -87,10 +96,11 @@ def before_search(self) -> None:
             "Only `export_config` constraint is supported for pruning!"
         )
 
+        self.constraints["export_config"] = copy.deepcopy(self.constraints["export_config"])
         export_config = self.constraints["export_config"]
         assert isinstance(export_config, dict)  # to keep mypy happy
         assert export_config.keys() <= SUPPORTED_HPARAMS, (
-            f"Only {SUPPORTED_HPARAMS} are supported for pruning!"
+            f"Only {SUPPORTED_HPARAMS} are supported for pruning! Received: {export_config.keys()}"
         )
 
         assert ("num_attention_heads" in export_config and "num_query_groups" in export_config) or (
@@ -124,14 +134,37 @@ def before_search(self) -> None:
     def run_search(self) -> None:
         """Run actual search."""
         # Run forward loop to collect activations and sort parameters
-        assert self.forward_loop is not None
-        is_training = self.model.training
-        self.model.eval()
-        print_rank_0("Running forward loop...")
-        with torch.no_grad():
-            self.forward_loop(self.model)
-        sort_parameters(self.model, self.hps_to_sort, verbose=True)
-        self.model.train(is_training)
+        unwrapped_model = self.model
+        for m in self.model.modules():
+            if isinstance(m, _DynamicMCoreLanguageModel):
+                unwrapped_model = m
+                break
+        assert isinstance(unwrapped_model, _DynamicMCoreLanguageModel), "Model not supported!"
+
+        if self.layer_scores and self.activations_per_rank:  # Available from checkpoint
+            print_rank_0("Loading activations and scores per rank from checkpoint...")
+            unwrapped_model.set_activations_and_layer_scores(
+                self.activations_per_rank, self.layer_scores
+            )
+        elif not self.config["skip_sorting"]:
+            print_rank_0("Running forward loop...")
+            assert self.forward_loop is not None
+            is_training = self.model.training
+            self.model.eval()
+            with torch.no_grad():
+                self.forward_loop(self.model)
+            self.model.train(is_training)
+
+            # Store activations and layer scores for re-pruning with different export configs
+            self.activations_per_rank, self.layer_scores = (
+                unwrapped_model.get_activations_and_layer_scores()
+            )
+            self.save_search_checkpoint(verbose=True)
+
+        if self.config["skip_sorting"]:
+            print_rank_0("Skipping sorting parameters...")
+        else:
+            sort_parameters(self.model, self.hps_to_sort, verbose=True)
 
         # Prune homogeneously
         export_config = self.constraints["export_config"]

modelopt/torch/utils/distributed.py

@@ -124,23 +124,23 @@ def broadcast(obj: Any, src: int = 0, group=None) -> Any:
     return obj
 
 
-def _allgather(tensors: list[torch.Tensor], tensor: torch.Tensor) -> None:
+def _allgather(tensors: list[torch.Tensor], tensor: torch.Tensor, group=None) -> None:
     if backend() == "torch":
-        torch.distributed.all_gather(tensors, tensor)
+        torch.distributed.all_gather(tensors, tensor, group)
 
 
-def allgather(obj: Any) -> list[Any]:
+def allgather(obj: Any, group=None) -> list[Any]:
     """Gathers an object from all processes into a list."""
-    if size() == 1:
+    if size(group) == 1:
         return [obj]
 
     # serialize
     tensor = _serialize(obj).cuda()
 
     # gather the tensor size
     tensor_size = torch.LongTensor([tensor.numel()]).cuda()
-    tensor_sizes = [torch.LongTensor([0]).cuda() for _ in range(size())]
-    _allgather(tensor_sizes, tensor_size)
+    tensor_sizes = [torch.LongTensor([0]).cuda() for _ in range(size(group))]
+    _allgather(tensor_sizes, tensor_size, group)
     tensor_sizes = [int(tensor_size.item()) for tensor_size in tensor_sizes]
     max_size = max(tensor_sizes)
 
@@ -149,7 +149,7 @@ def allgather(obj: Any) -> list[Any]:
     if tensor_size != max_size:
         padding = torch.ByteTensor(size=(max_size - tensor_size,)).cuda()
         tensor = torch.cat((tensor, padding), dim=0)
-    _allgather(tensors, tensor)
+    _allgather(tensors, tensor, group)
 
     # deserialize
     objs = []
@@ -159,9 +159,9 @@ def allgather(obj: Any) -> list[Any]:
     return objs
 
 
-def allreduce(obj: Any, reduction: str = "sum") -> Any:
+def allreduce(obj: Any, reduction: str = "sum", group=None) -> Any:
     """Reduces an object from all processes."""
-    objs = allgather(obj)
+    objs = allgather(obj, group)
     if reduction == "sum":
         return sum(objs)
     else: