NVIDIA
diff --git a/‎CHANGELOG.rst‎
Lines changed: 3 additions & 1 deletion b/‎CHANGELOG.rst‎
Lines changed: 3 additions & 1 deletion
diff --git a/‎docs/source/guides/3_pruning.rst‎
Lines changed: 3 additions & 2 deletions b/‎docs/source/guides/3_pruning.rst‎
Lines changed: 3 additions & 2 deletions
diff --git a/‎examples/hf_ptq/hf_ptq.py‎
Lines changed: 0 additions & 1 deletion b/‎examples/hf_ptq/hf_ptq.py‎
Lines changed: 0 additions & 1 deletion
diff --git a/‎examples/megatron_bridge/README.md‎
Lines changed: 27 additions & 1 deletion b/‎examples/megatron_bridge/README.md‎
Lines changed: 27 additions & 1 deletion
diff --git a/‎examples/megatron_bridge/export.py‎
Lines changed: 3 additions & 0 deletions b/‎examples/megatron_bridge/export.py‎
Lines changed: 3 additions & 0 deletions
@@ -21,11 +21,14 @@ Changelog
 - Add the ``day0-release`` agent skill (``.agents/skills/day0-release/``), a deterministic end-to-end driver that chains the PTQ → evaluation → comparison skills (the evaluation stage deploys the checkpoint itself) with an enforced gate after each stage and returns a publish decision (ACCEPT / REGRESSION / ANOMALOUS / INFEASIBLE). Ships three GPU-free, unit-tested gate scripts (``gate_ptq.py``, ``gate_run.py``, ``gate_compare.py``) that validate checkpoint coverage, evaluation-run completeness, and baseline-vs-candidate accuracy threshold. v1 reports and stops on regression; the recipe-search loop is deferred.
 - Add **streaming** speculative-decoding training (EAGLE3 / DFlash): the draft trains on base-model hidden states produced on the fly by a co-located ``vllm serve`` (no disk dump), moved trainer-side over NIXL RDMA, scaling to multi-node (dedicated serve replicas + DDP trainers). New launcher examples for NVFP4 Kimi-K2.5 / K2.6 on GB200/aarch64 under ``tools/launcher/examples/moonshotai/``.
 - Add a fused Triton fast path for ``local_hessian`` NVFP4 weight-scale search (the Hessian-weighted FP8-E4M3 scale sweep). For each NVFP4 block it minimizes ``dwᵀ H dw`` over the 126 candidate scales using the per-cin-block local Hessian on tensor cores, replacing the per-weight Python reference sweep — roughly **34x** faster on a single 8192x4096 weight and bit-exact with the reference for fp32/fp16 weights. Used automatically during ``local_hessian`` calibration for both dense and fused-MoE expert weights; falls back to the reference sweep on CPU, when Triton is unavailable, or via ``MODELOPT_NVFP4_TRITON_SWEEP=0``.
+- Add **context-parallel (CP)** and **data-parallel (DP)** support to the shared Megatron-Core inference/calibration utilities. Under CP, ``get_megatron_calibration_forward_loop`` and ``megatron_mmlu`` partition each sequence across CP ranks (zigzag load-balanced), ``megatron_prefill`` accepts a CP-partitioned ``position_ids`` and lets the CP-aware causal attention build the mask, and MMLU gathers per-rank logits back to the full sequence for last-token scoring. Under DP, calibration shards the dataset across data-parallel ranks (``DistributedSampler``; amax is max-reduced across the DP group inside ``mtq``) and ``megatron_mmlu`` shards whole batches across DP ranks and all-reduces the per-subject counts. DP is implicit (``world_size / (tp * pp * cp)``); ``examples/megatron_bridge/quantize.py`` gains a ``--cp_size`` flag.
 - Add Minitron pruning support for Megatron-Core models with the following new attention and MoE variants. For these, only ``hidden_size`` is pruned (alongside the usual ``ffn_hidden_size`` / ``num_layers`` / MoE dimensions); the variant-internal dimensions noted below are not pruned:
 
   - **GatedDeltaNet** (linear attention) and **gated attention** (``attention_output_gate``), such as Qwen3.5 (hybrid GatedDeltaNet + gated-attention) language models, including MoE variants — attention / linear-attention heads are not pruned.
   - **Multi-Latent Attention (MLA)**, such as DeepSeek — MLA latent ranks are not pruned.
   - **Latent MoE**, such as Nemotron-3-Super — ``hidden_size`` pruning resizes the latent projections while the MoE latent dim itself is not pruned.
+- Add Minitron pruning support for the language model part of vision-language models (e.g. Qwen3.5-VL, Gemma3-VL) via ``examples/megatron_bridge/prune_minitron.py``. The language model is pruned while the vision tower is left intact and the full VLM is saved back; ``hidden_size`` is not pruned if it is shared with the vision projector. Pruning importance is estimated from image-text calibration (the full VLM forward over vision-conditioned activations) by default, or from a text dataset for text-only ablations.
+- Add PTQ support for the language model part of vision-language models (e.g. Qwen3.5-VL, Gemma3-VL) via ``examples/megatron_bridge/quantize.py``. Only the language model is quantized (vision tower + projector left in full precision) and the full VLM is saved as a Megatron checkpoint. The calibration modality is inferred from ``--calib_dataset_name``: an image-text dataset drives the full VLM forward (vision-conditioned activations), while a text dataset runs text-only calibration of the language model. Image-text calibration shards across data-parallel ranks (context parallelism is supported only for text-only calibration). HuggingFace unified export of a quantized VLM is not yet supported.
 - Add NVFP4 Four-Over-Six (4/6) weight quantization (``mtq.NVFP4_FOUR_OVER_SIX_CFG``): MSE weight calibration picks, per block, between an M=6 and an M=4 dynamic range (the choice is folded into the FP8 per-block scales), with the ``four_over_six: true`` flag normalizing those scales by 256 (vs 448) for M=4 headroom. Supported via ``mtq.quantize`` and HF / Megatron export only -- **not** ``mtq.compress``, which does not preserve the per-block M=4/M=6 choice
 - Add dLLM (tied-weight PTQ and HF-checkpoint export) support for diffusion-based encoder-decoder LLMs (e.g. DiffusionGemma) whose encoder/decoder stacks share parameters via HF ``_tied_weights_keys``.
 
@@ -34,7 +37,6 @@ Changelog
   - The exported state_dict is also **reordered (decoder keys win instead of encoder)** so canonical-side keys per HF's ``_tied_weights_keys`` declaration win the data_ptr dedup; gated to the DiffusionGemma model class in ``_reorder_canonical_first``, no-op for every other model.
   - New DiffusionGemma model-specific recipe under ``modelopt_recipes/huggingface/diffusion_gemma/ptq/`` (``nvfp4_experts_only.yaml`` + its ``disabled_quantizers.yaml`` unit) adds the ``*self_conditioning*`` exclude on top of the standard default, leaving the shared ``default_disabled_quantizers`` unit clean for non-diffusion models — pattern matches the existing ``phi4mm`` / ``nemotron_vl`` model-specific recipes.
   - ``hf_ptq.py`` also unwraps ``ModelOutput`` dataclasses from ``.generate()`` so the preview decode works on diffusion models. Non-tied models see no behavioral change.
-- Add **context-parallel (CP)** and **data-parallel (DP)** support to the shared Megatron-Core inference/calibration utilities. Under CP, ``get_megatron_calibration_forward_loop`` and ``megatron_mmlu`` partition each sequence across CP ranks (zigzag load-balanced), ``megatron_prefill`` accepts a CP-partitioned ``position_ids`` and lets the CP-aware causal attention build the mask, and MMLU gathers per-rank logits back to the full sequence for last-token scoring. Under DP, calibration shards the dataset across data-parallel ranks (``DistributedSampler``; amax is max-reduced across the DP group inside ``mtq``) and ``megatron_mmlu`` shards whole batches across DP ranks and all-reduces the per-subject counts. DP is implicit (``world_size / (tp * pp * cp)``); ``examples/megatron_bridge/quantize.py`` gains a ``--cp_size`` flag.
 - Add **Domino** speculative-decoding training: the parallel DFlash draft backbone plus a lightweight GRU causal correction head, selected via ``dflash_architecture_config.projector_type=domino``. Trained with a base/final dual loss whose ``dflash_lambda_base_start``/``dflash_lambda_base_decay_ratio`` curriculum decays the base-loss weight 1→0. Exports in the z-lab drafter format; recipe at ``modelopt_recipes/general/speculative_decoding/domino.yaml``. Training only — the inference path is not wired up yet.
 
 **Bug Fixes**
 
@@ -16,9 +16,10 @@ These pruning methods support pruning the convolutional and linear layers, and
 attention heads of the model. More details on these pruning modes is as follows:
 
 #.  ``mcore_minitron``: A pruning method developed by NVIDIA Research for pruning GPT, Mamba and Hybrid
-    Transformer Mamba models in NVIDIA Megatron-Bridge or Megatron-LM framework. It uses the activation magnitudes to prune
+    Transformer Mamba models (including MoE, and the language model of vision-language models) in NVIDIA
+    Megatron-Bridge or Megatron-LM framework. It uses the activation magnitudes to prune
     the embedding hidden size, mlp ffn hidden size, transformer attention heads, GQA query groups,
-    mamba heads and head dimension, and number of layers of the model.
+    mamba heads and head dimension, MoE experts, and number of layers of the model.
     Checkout more details of the algorithm in the `paper <https://arxiv.org/abs/2408.11796>`_.
 #.  ``puzzletron``: An advanced LLM/VLM pruning method by NVIDIA using Mixed Integer Programming (MIP) based NAS search algorithm.
 #.  ``fastnas``: A pruning method recommended for Computer Vision models. Given a pretrained model,
 
@@ -266,7 +266,6 @@ def make_calib_dataloader(
             batch_size=args.batch_size,
             num_samples=args.calib_size[0],
             device=device,
-            max_length=args.calib_seq,
             require_image=True,
             subsets=["sparsetables", "plotqa_cot", "wiki_en"],
             shuffle_buffer_size=10_000,
 
@@ -96,7 +96,15 @@ torchrun --nproc_per_node 2 export.py \
 
 To see the full usage for advanced configurations, run `torchrun --nproc_per_node 1 quantize.py --help` (or `export.py --help`).
 
-For VLM (vision-language model) quantization, see the Megatron-Bridge repository [here](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/examples/quantization).
+### Vision-Language Models (VLMs)
+
+For a vision-language model (e.g. Qwen3.5-VL, Gemma3-VL), `quantize.py` automatically quantizes only the **language model** and leaves the vision tower and vision-language projector in full precision, then saves the full VLM back as a Megatron checkpoint. The calibration modality is inferred from `--calib_dataset_name`:
+
+- An **image-text** dataset (the default for VLMs, `nemotron_vlm_dataset_v2`) drives the full VLM forward, so the language model is calibrated on vision-conditioned activations.
+- A **text** dataset runs text-only calibration of the language model (vision tower idle).
+
+> [!NOTE]
+> HuggingFace unified export (`export.py`) of a quantized VLM is not yet supported; the quantized VLM is saved in Megatron checkpoint format only.
 
 ## Sanity-Check Generation
 
@@ -307,9 +315,27 @@ torchrun --nproc_per_node 1 prune_minitron.py --help
 > [!NOTE]
 > NAS-based pruning requires ~2x the GPU memory of Manual pruning because it needs to simultaneously hold original model while evaluating each pruned candidate.
 
+> [!NOTE]
+> Multi-token-prediction (MTP) heads (e.g. Qwen3.5) are not pruned yet — they are dropped for the prune run and the saved checkpoint has no MTP. Autoregressive inference is unaffected; for speculative decoding, run a short MTP SFT on the pruned model.
+
 > [!NOTE]
 > If pruning a Nemotron model and you want to save the pruned model back in HF format, please downgrade to `transformers<5` via `python -m pip install "transformers<5"` before pruning.
 
+### Vision-Language Models (VLMs)
+
+For a vision-language model (e.g. Qwen3.5-VL, Gemma3-VL), `prune_minitron.py` automatically prunes only the **language model** and leaves the vision tower intact, then saves the full VLM back. All the pruning modes above (parameter count, active parameter count, memory footprint, and manual `export_config`) work unchanged, with two VLM-specific caveats:
+
+- The `--prune_target_params` / `--prune_target_active_params` / `--prune_target_memory_mb` targets (and `export_config` dimensions) apply to the **language model only** — the (unpruned) vision tower's parameters are *not* counted, so the full saved VLM will be larger than the target.
+- `hidden_size` is never pruned for VLMs (it is shared with the vision projector).
+
+```bash
+torchrun --nproc_per_node 2 prune_minitron.py \
+    --pp_size 2 \
+    --hf_model_name_or_path Qwen/Qwen3.5-4B \
+    --prune_target_params 3e9 \
+    --output_hf_path /tmp/Qwen3.5-4B-Pruned-3B
+```
+
 ## Resources
 
 - 📅 [Roadmap](https://github.com/NVIDIA/Model-Optimizer/issues/1699)
 
@@ -144,6 +144,9 @@ def main(args: argparse.Namespace):
     print_rank_0(
         f"Exporting to HuggingFace (unified) checkpoint at {args.export_unified_hf_path}..."
     )
+    # TODO (OMNIML-5366): quantized-VLM HF export. export_mcore_gpt_to_hf's per-arch mappings don't
+    # cover Qwen3.5-VL / Gemma3-VL; See if Megatron-Bridge's AutoBridge.export_hf_weights_quant can be
+    # used instead.
     export_mcore_gpt_to_hf(
         unwrapped_model,
         args.hf_model_name_or_path,