feat(speculative): add EAGLE-3 sequence packing and reasoning-mode control by khazic · Pull Request #2444 · NVIDIA-NeMo/Automodel

khazic · 2026-06-07T08:25:46Z

What does this PR do ?

Adds two EAGLE-3 training features for the speculative-decoding stack:

Sequence packing (packed_sequence_size > 0, colocated target backend
only). Variable-length chat samples are greedily packed into fixed-width
rows, removing the per-sample padding waste of the default max_length
path. Documents within a row attend block-causally: the target uses a 4D
block-causal mask under SDPA, the draft uses varlen FlashAttention-2, and
cross-document TTT supervision is gated by doc_remaining so deeper steps
never leak across document boundaries.
Reasoning-mode control. A --reasoning {none,save,disable} flag on
regenerate.py and a mask_reasoning_content option on the
EAGLE/P-EAGLE/DFlash recipes, to keep or drop target reasoning traces from
the loss mask.

Both features are opt-in and leave the existing default paths unchanged.

Changelog

Add build_packed_eagle3_dataset + _pack_collate and a
packed_sequence_size knob to the EAGLE-3 dataloader.
Add build_block_causal_additive_mask (block-causal 4D additive mask) to
packed_sequence.py; no existing function is modified.
Add varlen FlashAttention-2 Block-1 path (_flash_block1_varlen,
_seq_lens_to_cu_seqlens) to the EAGLE-3 draft attention.
Thread position_ids / seq_lens / doc_remaining through the target
wrapper and the trainer module; gate cross-document supervision per TTT step.
Wire packed_sequence_size into train_eagle3.py (colocated backend only;
remote raises NotImplementedError).
Add --reasoning to regenerate.py and mask_reasoning_content to the
EAGLE/P-EAGLE/DFlash recipes and example configs.
Tests: test_eagle3_packing.py, test_eagle3_packing_fa2_parity.py
(FA2-vs-eager / packed-vs-padded parity + speedup), and test_regenerate.py.

Before your PR is "Ready for review"

Pre checks:

Make sure you read and followed Contributor guidelines
Did you write any new necessary tests?
Did you add or update any necessary documentation?

Additional Information

GPU parity validation (CUDA + flash-attn 2.8.3, Qwen3-class Llama configs):

test_eagle3_flash_attention.py (draft FA2 vs eager, non-packed): both cases
match on valid tokens; left-padding is correctly rejected.

test_eagle3_packing_fa2_parity.py (packed varlen-FA2 vs padded reference,
88 documents, micro_batch_size=1):

valid supervised toks : 1312976 (both layouts)
loss  no-packing      : 7.726586
loss  packing         : 7.726586
loss  |abs| / rel     : 2.529e-08 / 3.273e-09
max grad rel diff     : 6.204e-05  (draft_model.model.embed_tokens.weight)
full-pass no-packing  : 18227.67 ms (88 rows x 4096 = 360448 tok)
full-pass packing     : 10778.84 ms (52 rows x 4096 = 212992 tok)
speedup (padded/packed): 1.69x

Loss is parity to ~3e-9 relative and gradients to ~6e-5 relative, with a 1.69x
step speedup from eliminating padding waste.

…h training Add --reasoning {none,save,disable} flag to regenerate.py for controlling whether target model reasoning content is preserved or suppressed during data regeneration. Add mask_reasoning_content option to EAGLE/P-EAGLE/DFlash training recipes to exclude reasoning traces from the loss mask. Co-authored-by: khazic <khazzz1c@gmail.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com>

copy-pr-bot · 2026-06-07T08:25:49Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

HuiyingLi · 2026-06-07T08:28:34Z

/ok to test 14a8e7f

Pack variable-length chat samples into fixed-width rows for EAGLE-3 training, removing the per-sample padding waste of the default max_length path. Documents within a row attend block-causally: the target uses a 4D block-causal mask (SDPA) and the draft uses varlen FlashAttention-2; cross-document TTT supervision is gated by doc_remaining so deeper steps never leak across boundaries. Opt-in via packed_sequence_size > 0, colocated target backend only. Covered by unit tests plus an FA2-vs-eager parity test. Co-authored-by: khazic <khazzz1c@gmail.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com>

HuiyingLi · 2026-06-07T08:47:34Z

/ok to test 160078c

HuiyingLi · 2026-06-08T03:48:27Z

/ok to test 6ce90d1

…ntrol (#2444) * feat(speculative): add reasoning mode control for EAGLE/P-EAGLE/DFlash training Add --reasoning {none,save,disable} flag to regenerate.py for controlling whether target model reasoning content is preserved or suppressed during data regeneration. Add mask_reasoning_content option to EAGLE/P-EAGLE/DFlash training recipes to exclude reasoning traces from the loss mask. Co-authored-by: khazic <khazzz1c@gmail.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> * feat(speculative): add EAGLE-3 sequence packing for draft training Pack variable-length chat samples into fixed-width rows for EAGLE-3 training, removing the per-sample padding waste of the default max_length path. Documents within a row attend block-causally: the target uses a 4D block-causal mask (SDPA) and the draft uses varlen FlashAttention-2; cross-document TTT supervision is gated by doc_remaining so deeper steps never leak across boundaries. Opt-in via packed_sequence_size > 0, colocated target backend only. Covered by unit tests plus an FA2-vs-eager parity test. Co-authored-by: khazic <khazzz1c@gmail.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> --------- Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> Co-authored-by: thyways <2484113689@qq.com> Co-authored-by: Huiying <willwin.lee@gmail.com>

* make mesh accept meshcontext Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(transformers): resolve mesh context inputs Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * rm Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * use moe_overrides Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * make create_mesh_context the entry point for dist setup Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix: add renamed distributed utility files Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(vlm): complete dist_setup -> mesh_context rename Two leftover references to the old setup_distributed/dist_setup API were missed when the recipe was migrated to create_mesh_context_from_config: - nemo_automodel/recipes/vlm/finetune.py:794 still read self.dist_setup.cp_size, which would AttributeError on any PP+CP VLM run. - tests/unit_tests/recipes/test_finetune_vlm_cp_wiring.py monkeypatched the stale symbol "setup_distributed", causing three parametrizations of test_setup_skips_pp_media_prechunk_when_cp_preembeds_vlm_inputs to fail during pytest setup with AttributeError. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * remove activation checkpointing from meshcontext Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * refac Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * dedup Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * Update nemo_automodel/_diffusers/auto_diffusion_pipeline.py Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * Update skills/distributed-training/SKILL.md Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * Update skills/distributed-training/SKILL.md Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * Update nemo_automodel/components/distributed/config.py Co-authored-by: Adil <47084919+adil-a@users.noreply.github.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * Update skills/nemo-automodel-distributed-training/SKILL.md Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * docs(distributed): update setup API example Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * Update nemo_automodel/_diffusers/auto_diffusion_pipeline.py Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(diffusers): remove duplicated DistributedSetup.build call in fsdp2 path Commit 6acf01f left a duplicated `distributed_setup = DistributedSetup.build(` line and a stray extra `)` in the fsdp2 branch of the parallelism-manager builder. This broke `ruff format --check` (CI linting job) and, once formatted, would have nested the call as `DistributedSetup.build(distributed_setup=...)` — a kwarg the builder does not accept. Restore the single build() call matching the ddp branch. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(vlm): pass distributed_setup through Gemma4 joint drafter composite The distributed-config API refactor routes all distributed settings through a single ``distributed_setup`` argument and rejects the separate ``moe_mesh`` / ``distributed_config`` / ``pipeline_config`` kwargs in ``NeMoAutoModel*.from_pretrained``. ``Gemma4WithDrafter.from_pretrained`` was still forwarding those separate kwargs to its inner base/drafter loads, so the joint-drafter VLM finetune (L2_HF_Transformer_VLM_Gemma4_Joint_Drafter) raised ``TypeError: Distributed settings must be passed with distributed_setup``. Forward ``distributed_setup`` to both sub-loaders instead, and extend the pipeline/context-parallel safety guards to read pp_size/cp_size from the resolved setup so the KV-sharing invariant holds on the recipe path too. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Adil Asif <adasif@nvidia.com> * feat: add use_memory_efficient_lora knob (#2239) * add use_memory_efficient_lora knob Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * add use_memory_efficient_lora Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * delete sft gpt-oss 20b single gpu Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * add nemotron nano v3 single-gpu lora example Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * add grad ckpt Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * add fused lora mlp Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(checkpoint): support single-GPU Nemotron-H MoE LoRA checkpoint load Loading a merged-expert Nemotron-H MoE checkpoint through the default DCP / set_model_state_dict path transiently materializes a second on-device copy of the expert weights, which OOMs a 30B-class model on a single 80GB GPU. Checkpointer.load now routes single-device custom-model safetensors through the frugal full-state path (load to CPU, merge from_hf on CPU, copy into the model), keeping device memory at ~model size. _load_full_state_dict_into_model normalizes stray real (CPU) buffers left behind by custom-model meta materialization onto the parameter device (avoids 'Multiple devices found'), and uses plain load_state_dict for non-DTensor models so the full state dict is not moved on-device a second time. Adds a [nemotron-singlegpu-lora] note plus per-site tags documenting these single-device special cases, links the exercising recipe (examples/llm_finetune/nemotron/nemotron_nano_v3_singlegpu_lora.yaml), and flags the load path for a future refactor. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * feat(peft): add fused LoRA SwiGLU/ReLU² MLP with recompute backward Fuses gate+up+down+activation into a single autograd Function that saves only (x, gate_out, up_out) and recomputes the activation and down-projection input in backward, roughly halving MLP activation memory at equal speed during LoRA SFT. SwiGLU forward/backward use elementwise Triton kernels (with in-place backward buffer reuse) and a pure-torch fallback when Triton is unavailable; matmuls stay on cuBLAS. Covers SiLU-SwiGLU (gate/up/down) and non-gated ReLU² (e.g. Nemotron-H dense) MLPs. install_fused_lora_mlp() swaps each LoRA-applied MLP's forward and falls back to the per-linear path at runtime under DTensor (TP/EP), DoRA, or active dropout, keeping it correct under sharding. Already wired from lora.py; opt out via NEMO_AUTOMODEL_DISABLE_FUSED_LORA_MLP=1. Activation recompute follows Megatron-Core's SwiGLUFunction; the fused LoRA-MLP and in-place buffer reuse follow Unsloth's LoRA_MLP (both Apache-2.0). Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * refactor(peft): drop NEMO_AUTOMODEL_DISABLE_FUSED_LORA_MLP env knob The fused LoRA MLP can already be disabled via the use_memory_efficient_lora config flag, and fusion auto-falls-back per-MLP under DTensor / DoRA / active dropout. The env var was a redundant escape hatch; remove it and the now-unused os import. Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * test(checkpoint): align custom-model load-routing guard with single-device fast path The nemotron-singlegpu-lora change routes single-device (world_size == 1) custom safetensors models through the frugal full-state fast path instead of DCP. The fast path now applies the state_dict_adapter from_hf conversion on CPU (_maybe_adapt_state_dict_from_hf), so custom MoE expert merging still happens — the guard test's original premise (fast path bypasses conversion) no longer holds. - Reframe test_custom_model_skips_fast_path_uses_dcp as the multi-rank (sharded) case (WORLD_SIZE=2), where DCP per-rank DTensor slicing is genuinely required. - Add test_single_device_custom_model_uses_fast_path covering the new world_size==1 behavior (fast path used, DCP not). Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> --------- Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * fix(deepseek_v3): initialize weights in fp32 and default router to fp32 (#2450) * fix(deepseek_v3): init weights in fp32, default router to fp32 Sampling the random init directly in bf16 distorts the variance/mean schedule and produces exploding first-step gradients (flat/diverging loss) for from-scratch pretraining. Add an init_weights_in_fp32 context manager that samples in fp32 and casts back to the resident dtype, and use it in DeepSeek-V3 initialize_weights. Also default the router (gate_precision) to fp32 to match the HF reference. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * refactor(models): rename init_weights_in_fp32 to yield_fp32_model Generalize the context manager per review: it's a generic "run this block with the model in fp32" tool, not init-specific. Yield the model and make the exit dtype optional (defaults to the model's pre-context float dtype). Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> --------- Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * fix(multimodal): migrate finetune recipe to DistributedSetup/MeshContext API The auto-class-public-api refactor deleted `recipes/_dist_setup.py` and moved recipes to the `DistributedSetup` / `MeshContext` API, but `multimodal/finetune.py` was left importing and calling the deleted `_dist_setup.setup_distributed`, so importing the module raised ModuleNotFoundError and broke the import-check in every Pip/UV install job. Migrate it to the shared `_distributed_setup_attributes(create_distributed_setup_from_config(...))` pattern used by the llm/vlm recipes: unpack distributed_setup / mesh_context / distributed_config / device_mesh / moe_mesh / pp_enabled / pipeline_config / moe_parallel_config / activation_checkpointing, and update the model-build calls (`mesh=self.mesh_context`, `moe_config`/`cfg_moe=self.moe_parallel_config`, `activation_checkpointing=self.activation_checkpointing`). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> * feat(speculative): add EAGLE-3 sequence packing and reasoning-mode control (#2444) * feat(speculative): add reasoning mode control for EAGLE/P-EAGLE/DFlash training Add --reasoning {none,save,disable} flag to regenerate.py for controlling whether target model reasoning content is preserved or suppressed during data regeneration. Add mask_reasoning_content option to EAGLE/P-EAGLE/DFlash training recipes to exclude reasoning traces from the loss mask. Co-authored-by: khazic <khazzz1c@gmail.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> * feat(speculative): add EAGLE-3 sequence packing for draft training Pack variable-length chat samples into fixed-width rows for EAGLE-3 training, removing the per-sample padding waste of the default max_length path. Documents within a row attend block-causally: the target uses a 4D block-causal mask (SDPA) and the draft uses varlen FlashAttention-2; cross-document TTT supervision is gated by doc_remaining so deeper steps never leak across boundaries. Opt-in via packed_sequence_size > 0, colocated target backend only. Covered by unit tests plus an FA2-vs-eager parity test. Co-authored-by: khazic <khazzz1c@gmail.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> --------- Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> Co-authored-by: thyways <2484113689@qq.com> Co-authored-by: Huiying <willwin.lee@gmail.com> * feat(distributed): add selective activation checkpointing for FSDP2 (#2389) * feat(distributed): add selective activation checkpointing for FSDP2 Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * fix(distributed): support selective activation checkpointing with torch.compile Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs(fern): drop selective AC from frozen v0.4 snapshot Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * feat(distributed): honor selective activation checkpointing on single GPU Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * feat(moe): support selective activation checkpointing with expert parallelism Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * fix(model): make DeepSeek MLP dispatch wrapper-safe Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * fix(distributed): save expert grouped-GEMM in selective AC and add op trace Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * feat(moe): compile selective activation checkpointing wrappers outer Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * refactor(distributed): move selective AC into its own module Extract the TorchTitan-style selective activation checkpointing core out of the central parallelizer.py into a dedicated activation_checkpointing.py: op-set construction, the save/recompute policy, block/sub-module wrappers, KV-sharing detection, and the compile-outer wrapper flag. parallelizer.py keeps only the thin apply_selective_activation_checkpointing entry point, which still needs the heavy, transformers-aware _extract_model_layers, so the dependency stays one-directional (parallelizer -> activation_checkpointing -> parallelizer_utils) with no circular imports. Move the opt-in NEMO_SELECTIVE_AC_TRACE diagnostic out of parallelizer.py into parallelizer_utils.maybe_trace_selective_ac_decision so the hot policy is a single call site instead of trace globals plus a helper. Make the new module's cross-module interface public (drop the leading underscore) and keep internal op-resolution/plumbing private. Update the moe and fsdp2 consumers and the unit tests to import from the new module. Also fix doc wording: clarify that torch.compile must be held fixed when comparing full vs. selective, and refer to TorchTitan as a reference implementation rather than "upstream". Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * refactor(distributed): move selective-AC trace into the AC module Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * test(distributed): patch activation_checkpointing.checkpoint_wrapper after AC module split Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs: apply tech-writer edits to gradient-checkpointing guide Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> --------- Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * feat(diffusion): improve qwen image finetuning configs (#2442) Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com> Co-authored-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * ci: add nemo-run, split qwen-vl-utils from decord for arm (#2456) * ci: add nemo-run, split qwen-vl-utils from decord for arm Signed-off-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * fix: override in pytorch container Signed-off-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> * Update uv lock Signed-off-by: NeMo Bot <nemo-bot@nvidia.com> --------- Signed-off-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Signed-off-by: NeMo Bot <nemo-bot@nvidia.com> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com> Co-authored-by: NeMo Bot <nemo-bot@nvidia.com> * Apply suggestions from code review Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * fix(precision): dtype contract bug fixes for FSDP2 mixed-dtype loads (#2419) * fix(transformers): unify loaded HF dtype via promote_types Make _restore_loaded_model_dtype dtype-aware: instead of always restoring to the checkpoint dtype, unify each floating tensor to promote_types(checkpoint, requested). This honors an explicit fp32 request while preserving intrinsically-fp32 checkpoint params (e.g. A_log) under a bf16 request, and is a no-op for the bf16/auto path. Fixes FSDP2 uniform-dtype tripping on HF mixed-dtype loads. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * feat(distributed): default pipeline dtype to FSDP activation dtype When pipeline parallelism is enabled and pipeline.dtype is unset, derive it from the FSDP mixed-precision activation dtype (mp_policy.output_dtype, falling back to param_dtype) so pipeline stage shape inference matches the real activation dtype (e.g. bf16 compute under fp32 master weights). An explicitly set pipeline.dtype is honored but warned on mismatch, since it can corrupt inter-stage recv buffers. No-ops for strategies without an mp_policy (e.g. MegatronFSDP) and for pp_size==1. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> (cherry picked from commit 3f6b246) * refactor(distributed): resolve FSDP compute dtype per-param, decoupled from storage fully_shard_by_dtype now groups parameters by their required *compute* dtype instead of their storage dtype, so fp32 master weights (uniform fp32 storage) still compute the bulk in mp_policy.param_dtype (bf16) while intrinsically-fp32 params keep fp32 compute. Per-parameter compute dtype is resolved by precedence: pinned fp32 (_keep_in_fp32_modules_strict) > HF-recorded checkpoint dtype (tagged onto each tensor at load time in _restore_loaded_model_dtype) > mp_policy.param_dtype. Qwen3.5's GatedDeltaNet fp32 holder is declared via patch_hf_model; the NemotronH and Qwen3.5 strategies thread the declaration through. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> (cherry picked from commit 3dd6b97) * docs(model-onboarding): document _keep_in_fp32_modules_strict contract Add SKILL.md §2.6 explaining which params must compute in fp32 (SSM A_log/ dt_bias/D, MoE sigmoid-gate bias, attention-sink bias, scale), how to declare them (class attribute vs patch_hf_model instance attribute), and why the pin is the robust signal across all load paths. Broaden the MoE checklist item and code comment accordingly. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> (cherry picked from commit a11db38) * test(distributed): add fp32 compute-dtype contract test Assert the resident compute dtype of every trainable parameter across the model archetypes that use fully_shard_by_dtype (dense, Qwen3.5-style hybrid), covering the full precedence chain: pinned fp32 > HF-recorded dtype > mp_policy.param_dtype, under fp32 master weights and ordinary loads. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> (cherry picked from commit dc83926) * feat(model): cast frozen modules to compute dtype to avoid mismatch Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> (cherry picked from commit d321f5e) * refactor(gemma4): drop projector dtype hook now general frozen cast handles it Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> (cherry picked from commit 1bc67e2) * feat(training): add dormant resolve_storage_dtype helper Add resolve_storage_dtype() (and its unit tests) for defaulting model.torch_dtype to fp32 for full-parameter torch.optim training. Not yet wired into recipes here; the call sites are marked with breadcrumb comments and enabled in a follow-up PR, keeping this PR limited to dtype bug fixes with no behavior/memory change. Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * fix(model): cast frozen-module buffers and unsharded params to compute dtype Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs(infra): correct frozen-tower FSDP comment to match sharding reality Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs(mixed-precision): clarify TE vs torch AdamW memory and precision trade-offs Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs(mixed-precision): apply tech writer edits Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs(mixed-precision): drop unresolvable FSDP anchor Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> --------- Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> * docs(speculative): add subsystem README, fold in regeneration guide (#2448) Add examples/speculative/README.md covering the whole speculative-decoding draft-training subsystem: supported methods (EAGLE-1/2/3/3.1, P-EAGLE, DFlash), target-model registry coverage, compute backends (eager vs flash_attention_2, flex_attention/sdpa, fused Triton soft cross-entropy, d2t/t2d draft-vocab compression), target backends (co-located, remote, offline cache), serving and benchmarking, inference-engine compatibility, and a consolidated config reference. Fold the standalone regenerate_with_target.md into the README's data preparation section (full two-step flow, tuning table, pitfalls) and remove the separate file so there is a single entry point. Signed-off-by: khazic <khazzz1c@gmail.com> * feat(diffusion): add Wan2.2 T2V-A14B two-stage finetuning support (#2284) * feat(diffusion): add Wan2.2 T2V-A14B two-stage finetuning support Signed-off-by: linnan wang <linnanw@nvidia.com> * fix the memory management for training large 14B wan model * fix wan2.2 support * all good for wan2.2 * update Signed-off-by: linnan wang <linnanw@nvidia.com> * docs(fern): add Wan2.2 T2V-A14B model coverage and release log entry Signed-off-by: linnan wang <linnanw@nvidia.com> * fix anther round of code review * fix(diffusion): sort wan.py imports to satisfy CI isort (I001) Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(diffusion): load inference checkpoints to CPU to halve peak GPU memory Avoids doubling peak GPU memory (and a potential OOM in Wan2.2 two-stage inference) by loading EMA/consolidated state dicts with map_location="cpu"; load_state_dict copies into the already-on-device parameters. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Signed-off-by: linnan wang <linnanw@nvidia.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> * test: include find_unused_parameters in ddp manager args expectation The DDP strategy config exposes find_unused_parameters (default False), so _build_diffusion_parallel_manager_args returns it in the ddp branch. Update the test's expected dict to match, fixing the L0 unit test failure. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: HuiyingLi <willwin.lee@gmail.com> * fix(distributed): address Claude review comments - infrastructure.py: forward the model wrapper's mp_policy (from FSDP2Config) to the MoE expert parallelizer when MoEParallelizerConfig.mp_policy is unset, so a custom precision policy isn't silently dropped for EP models. - skills/nemo-automodel-distributed-training/SKILL.md: fix stale references — MeshContext no longer holds strategy_config/pipeline_config/moe_config and STRATEGY_MAP moved to _STRATEGY_MAP in config.py; MoEParallelizerConfig now lives in components/distributed/config.py. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Signed-off-by: HuiyingLi <willwin.lee@gmail.com> --------- Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> Signed-off-by: Adil Asif <adasif@nvidia.com> Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com> Signed-off-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Signed-off-by: NeMo Bot <nemo-bot@nvidia.com> Signed-off-by: linnan wang <linnanw@nvidia.com> Signed-off-by: HuiyingLi <willwin.lee@gmail.com> Co-authored-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> Co-authored-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Yuhe Zhang <yuhezhang.zju@gmail.com> Co-authored-by: khazzz1c <khazzz1c@gmail.com> Co-authored-by: thyways <2484113689@qq.com> Co-authored-by: Huiying <willwin.lee@gmail.com> Co-authored-by: Pranav Thombre <pthombre@nvidia.com> Co-authored-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Co-authored-by: NeMo Bot <nemo-bot@nvidia.com> Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com> Co-authored-by: linnan wang <linnanw@nvidia.com>

feat: make mesh accept meshcontext (#2266) * make mesh accept meshcontext * fix(transformers): resolve mesh context inputs * rm * use moe_overrides * make create_mesh_context the entry point for dist setup * fix: add renamed distributed utility files * fix(vlm): complete dist_setup -> mesh_context rename Two leftover references to the old setup_distributed/dist_setup API were missed when the recipe was migrated to create_mesh_context_from_config: - nemo_automodel/recipes/vlm/finetune.py:794 still read self.dist_setup.cp_size, which would AttributeError on any PP+CP VLM run. - tests/unit_tests/recipes/test_finetune_vlm_cp_wiring.py monkeypatched the stale symbol "setup_distributed", causing three parametrizations of test_setup_skips_pp_media_prechunk_when_cp_preembeds_vlm_inputs to fail during pytest setup with AttributeError. * remove activation checkpointing from meshcontext * refac * dedup * fix * fix * Update nemo_automodel/_diffusers/auto_diffusion_pipeline.py * Update skills/distributed-training/SKILL.md * Update skills/distributed-training/SKILL.md * Update nemo_automodel/components/distributed/config.py * Update skills/nemo-automodel-distributed-training/SKILL.md * docs(distributed): update setup API example * Update nemo_automodel/_diffusers/auto_diffusion_pipeline.py * fix(diffusers): remove duplicated DistributedSetup.build call in fsdp2 path Commit 6acf01f left a duplicated `distributed_setup = DistributedSetup.build(` line and a stray extra `)` in the fsdp2 branch of the parallelism-manager builder. This broke `ruff format --check` (CI linting job) and, once formatted, would have nested the call as `DistributedSetup.build(distributed_setup=...)` — a kwarg the builder does not accept. Restore the single build() call matching the ddp branch. * fix(vlm): pass distributed_setup through Gemma4 joint drafter composite The distributed-config API refactor routes all distributed settings through a single ``distributed_setup`` argument and rejects the separate ``moe_mesh`` / ``distributed_config`` / ``pipeline_config`` kwargs in ``NeMoAutoModel*.from_pretrained``. ``Gemma4WithDrafter.from_pretrained`` was still forwarding those separate kwargs to its inner base/drafter loads, so the joint-drafter VLM finetune (L2_HF_Transformer_VLM_Gemma4_Joint_Drafter) raised ``TypeError: Distributed settings must be passed with distributed_setup``. Forward ``distributed_setup`` to both sub-loaders instead, and extend the pipeline/context-parallel safety guards to read pp_size/cp_size from the resolved setup so the KV-sharing invariant holds on the recipe path too. * feat: add use_memory_efficient_lora knob (#2239) * add use_memory_efficient_lora knob * add use_memory_efficient_lora * fix * delete sft gpt-oss 20b single gpu * add nemotron nano v3 single-gpu lora example * add grad ckpt * add fused lora mlp * fix(checkpoint): support single-GPU Nemotron-H MoE LoRA checkpoint load Loading a merged-expert Nemotron-H MoE checkpoint through the default DCP / set_model_state_dict path transiently materializes a second on-device copy of the expert weights, which OOMs a 30B-class model on a single 80GB GPU. Checkpointer.load now routes single-device custom-model safetensors through the frugal full-state path (load to CPU, merge from_hf on CPU, copy into the model), keeping device memory at ~model size. _load_full_state_dict_into_model normalizes stray real (CPU) buffers left behind by custom-model meta materialization onto the parameter device (avoids 'Multiple devices found'), and uses plain load_state_dict for non-DTensor models so the full state dict is not moved on-device a second time. Adds a [nemotron-singlegpu-lora] note plus per-site tags documenting these single-device special cases, links the exercising recipe (examples/llm_finetune/nemotron/nemotron_nano_v3_singlegpu_lora.yaml), and flags the load path for a future refactor. * feat(peft): add fused LoRA SwiGLU/ReLU² MLP with recompute backward Fuses gate+up+down+activation into a single autograd Function that saves only (x, gate_out, up_out) and recomputes the activation and down-projection input in backward, roughly halving MLP activation memory at equal speed during LoRA SFT. SwiGLU forward/backward use elementwise Triton kernels (with in-place backward buffer reuse) and a pure-torch fallback when Triton is unavailable; matmuls stay on cuBLAS. Covers SiLU-SwiGLU (gate/up/down) and non-gated ReLU² (e.g. Nemotron-H dense) MLPs. install_fused_lora_mlp() swaps each LoRA-applied MLP's forward and falls back to the per-linear path at runtime under DTensor (TP/EP), DoRA, or active dropout, keeping it correct under sharding. Already wired from lora.py; opt out via NEMO_AUTOMODEL_DISABLE_FUSED_LORA_MLP=1. Activation recompute follows Megatron-Core's SwiGLUFunction; the fused LoRA-MLP and in-place buffer reuse follow Unsloth's LoRA_MLP (both Apache-2.0). * refactor(peft): drop NEMO_AUTOMODEL_DISABLE_FUSED_LORA_MLP env knob The fused LoRA MLP can already be disabled via the use_memory_efficient_lora config flag, and fusion auto-falls-back per-MLP under DTensor / DoRA / active dropout. The env var was a redundant escape hatch; remove it and the now-unused os import. * test(checkpoint): align custom-model load-routing guard with single-device fast path The nemotron-singlegpu-lora change routes single-device (world_size == 1) custom safetensors models through the frugal full-state fast path instead of DCP. The fast path now applies the state_dict_adapter from_hf conversion on CPU (_maybe_adapt_state_dict_from_hf), so custom MoE expert merging still happens — the guard test's original premise (fast path bypasses conversion) no longer holds. - Reframe test_custom_model_skips_fast_path_uses_dcp as the multi-rank (sharded) case (WORLD_SIZE=2), where DCP per-rank DTensor slicing is genuinely required. - Add test_single_device_custom_model_uses_fast_path covering the new world_size==1 behavior (fast path used, DCP not). --------- * fix(deepseek_v3): initialize weights in fp32 and default router to fp32 (#2450) * fix(deepseek_v3): init weights in fp32, default router to fp32 Sampling the random init directly in bf16 distorts the variance/mean schedule and produces exploding first-step gradients (flat/diverging loss) for from-scratch pretraining. Add an init_weights_in_fp32 context manager that samples in fp32 and casts back to the resident dtype, and use it in DeepSeek-V3 initialize_weights. Also default the router (gate_precision) to fp32 to match the HF reference. * refactor(models): rename init_weights_in_fp32 to yield_fp32_model Generalize the context manager per review: it's a generic "run this block with the model in fp32" tool, not init-specific. Yield the model and make the exit dtype optional (defaults to the model's pre-context float dtype). --------- * fix(multimodal): migrate finetune recipe to DistributedSetup/MeshContext API The auto-class-public-api refactor deleted `recipes/_dist_setup.py` and moved recipes to the `DistributedSetup` / `MeshContext` API, but `multimodal/finetune.py` was left importing and calling the deleted `_dist_setup.setup_distributed`, so importing the module raised ModuleNotFoundError and broke the import-check in every Pip/UV install job. Migrate it to the shared `_distributed_setup_attributes(create_distributed_setup_from_config(...))` pattern used by the llm/vlm recipes: unpack distributed_setup / mesh_context / distributed_config / device_mesh / moe_mesh / pp_enabled / pipeline_config / moe_parallel_config / activation_checkpointing, and update the model-build calls (`mesh=self.mesh_context`, `moe_config`/`cfg_moe=self.moe_parallel_config`, `activation_checkpointing=self.activation_checkpointing`). * feat(speculative): add EAGLE-3 sequence packing and reasoning-mode control (#2444) * feat(speculative): add reasoning mode control for EAGLE/P-EAGLE/DFlash training Add --reasoning {none,save,disable} flag to regenerate.py for controlling whether target model reasoning content is preserved or suppressed during data regeneration. Add mask_reasoning_content option to EAGLE/P-EAGLE/DFlash training recipes to exclude reasoning traces from the loss mask. * feat(speculative): add EAGLE-3 sequence packing for draft training Pack variable-length chat samples into fixed-width rows for EAGLE-3 training, removing the per-sample padding waste of the default max_length path. Documents within a row attend block-causally: the target uses a 4D block-causal mask (SDPA) and the draft uses varlen FlashAttention-2; cross-document TTT supervision is gated by doc_remaining so deeper steps never leak across boundaries. Opt-in via packed_sequence_size > 0, colocated target backend only. Covered by unit tests plus an FA2-vs-eager parity test. --------- * feat(distributed): add selective activation checkpointing for FSDP2 (#2389) * feat(distributed): add selective activation checkpointing for FSDP2 * fix(distributed): support selective activation checkpointing with torch.compile * docs(fern): drop selective AC from frozen v0.4 snapshot * feat(distributed): honor selective activation checkpointing on single GPU * feat(moe): support selective activation checkpointing with expert parallelism * fix(model): make DeepSeek MLP dispatch wrapper-safe * fix(distributed): save expert grouped-GEMM in selective AC and add op trace * feat(moe): compile selective activation checkpointing wrappers outer * refactor(distributed): move selective AC into its own module Extract the TorchTitan-style selective activation checkpointing core out of the central parallelizer.py into a dedicated activation_checkpointing.py: op-set construction, the save/recompute policy, block/sub-module wrappers, KV-sharing detection, and the compile-outer wrapper flag. parallelizer.py keeps only the thin apply_selective_activation_checkpointing entry point, which still needs the heavy, transformers-aware _extract_model_layers, so the dependency stays one-directional (parallelizer -> activation_checkpointing -> parallelizer_utils) with no circular imports. Move the opt-in NEMO_SELECTIVE_AC_TRACE diagnostic out of parallelizer.py into parallelizer_utils.maybe_trace_selective_ac_decision so the hot policy is a single call site instead of trace globals plus a helper. Make the new module's cross-module interface public (drop the leading underscore) and keep internal op-resolution/plumbing private. Update the moe and fsdp2 consumers and the unit tests to import from the new module. Also fix doc wording: clarify that torch.compile must be held fixed when comparing full vs. selective, and refer to TorchTitan as a reference implementation rather than "upstream". * refactor(distributed): move selective-AC trace into the AC module * test(distributed): patch activation_checkpointing.checkpoint_wrapper after AC module split * docs: apply tech-writer edits to gradient-checkpointing guide --------- * feat(diffusion): improve qwen image finetuning configs (#2442) * ci: add nemo-run, split qwen-vl-utils from decord for arm (#2456) * ci: add nemo-run, split qwen-vl-utils from decord for arm * fix: override in pytorch container * Update uv lock --------- * Apply suggestions from code review * fix(precision): dtype contract bug fixes for FSDP2 mixed-dtype loads (#2419) * fix(transformers): unify loaded HF dtype via promote_types Make _restore_loaded_model_dtype dtype-aware: instead of always restoring to the checkpoint dtype, unify each floating tensor to promote_types(checkpoint, requested). This honors an explicit fp32 request while preserving intrinsically-fp32 checkpoint params (e.g. A_log) under a bf16 request, and is a no-op for the bf16/auto path. Fixes FSDP2 uniform-dtype tripping on HF mixed-dtype loads. * feat(distributed): default pipeline dtype to FSDP activation dtype When pipeline parallelism is enabled and pipeline.dtype is unset, derive it from the FSDP mixed-precision activation dtype (mp_policy.output_dtype, falling back to param_dtype) so pipeline stage shape inference matches the real activation dtype (e.g. bf16 compute under fp32 master weights). An explicitly set pipeline.dtype is honored but warned on mismatch, since it can corrupt inter-stage recv buffers. No-ops for strategies without an mp_policy (e.g. MegatronFSDP) and for pp_size==1. (cherry picked from commit 3f6b246) * refactor(distributed): resolve FSDP compute dtype per-param, decoupled from storage fully_shard_by_dtype now groups parameters by their required *compute* dtype instead of their storage dtype, so fp32 master weights (uniform fp32 storage) still compute the bulk in mp_policy.param_dtype (bf16) while intrinsically-fp32 params keep fp32 compute. Per-parameter compute dtype is resolved by precedence: pinned fp32 (_keep_in_fp32_modules_strict) > HF-recorded checkpoint dtype (tagged onto each tensor at load time in _restore_loaded_model_dtype) > mp_policy.param_dtype. Qwen3.5's GatedDeltaNet fp32 holder is declared via patch_hf_model; the NemotronH and Qwen3.5 strategies thread the declaration through. (cherry picked from commit 3dd6b97) * docs(model-onboarding): document _keep_in_fp32_modules_strict contract Add SKILL.md §2.6 explaining which params must compute in fp32 (SSM A_log/ dt_bias/D, MoE sigmoid-gate bias, attention-sink bias, scale), how to declare them (class attribute vs patch_hf_model instance attribute), and why the pin is the robust signal across all load paths. Broaden the MoE checklist item and code comment accordingly. (cherry picked from commit a11db38) * test(distributed): add fp32 compute-dtype contract test Assert the resident compute dtype of every trainable parameter across the model archetypes that use fully_shard_by_dtype (dense, Qwen3.5-style hybrid), covering the full precedence chain: pinned fp32 > HF-recorded dtype > mp_policy.param_dtype, under fp32 master weights and ordinary loads. (cherry picked from commit dc83926) * feat(model): cast frozen modules to compute dtype to avoid mismatch (cherry picked from commit d321f5e) * refactor(gemma4): drop projector dtype hook now general frozen cast handles it (cherry picked from commit 1bc67e2) * feat(training): add dormant resolve_storage_dtype helper Add resolve_storage_dtype() (and its unit tests) for defaulting model.torch_dtype to fp32 for full-parameter torch.optim training. Not yet wired into recipes here; the call sites are marked with breadcrumb comments and enabled in a follow-up PR, keeping this PR limited to dtype bug fixes with no behavior/memory change. * fix(model): cast frozen-module buffers and unsharded params to compute dtype * docs(infra): correct frozen-tower FSDP comment to match sharding reality * docs(mixed-precision): clarify TE vs torch AdamW memory and precision trade-offs * docs(mixed-precision): apply tech writer edits * docs(mixed-precision): drop unresolvable FSDP anchor --------- * docs(speculative): add subsystem README, fold in regeneration guide (#2448) Add examples/speculative/README.md covering the whole speculative-decoding draft-training subsystem: supported methods (EAGLE-1/2/3/3.1, P-EAGLE, DFlash), target-model registry coverage, compute backends (eager vs flash_attention_2, flex_attention/sdpa, fused Triton soft cross-entropy, d2t/t2d draft-vocab compression), target backends (co-located, remote, offline cache), serving and benchmarking, inference-engine compatibility, and a consolidated config reference. Fold the standalone regenerate_with_target.md into the README's data preparation section (full two-step flow, tuning table, pitfalls) and remove the separate file so there is a single entry point. * feat(diffusion): add Wan2.2 T2V-A14B two-stage finetuning support (#2284) * feat(diffusion): add Wan2.2 T2V-A14B two-stage finetuning support * fix the memory management for training large 14B wan model * fix wan2.2 support * all good for wan2.2 * update * docs(fern): add Wan2.2 T2V-A14B model coverage and release log entry * fix anther round of code review * fix(diffusion): sort wan.py imports to satisfy CI isort (I001) * fix(diffusion): load inference checkpoints to CPU to halve peak GPU memory Avoids doubling peak GPU memory (and a potential OOM in Wan2.2 two-stage inference) by loading EMA/consolidated state dicts with map_location="cpu"; load_state_dict copies into the already-on-device parameters. --------- * test: include find_unused_parameters in ddp manager args expectation The DDP strategy config exposes find_unused_parameters (default False), so _build_diffusion_parallel_manager_args returns it in the ddp branch. Update the test's expected dict to match, fixing the L0 unit test failure. * fix(distributed): address Claude review comments - infrastructure.py: forward the model wrapper's mp_policy (from FSDP2Config) to the MoE expert parallelizer when MoEParallelizerConfig.mp_policy is unset, so a custom precision policy isn't silently dropped for EP models. - skills/nemo-automodel-distributed-training/SKILL.md: fix stale references — MeshContext no longer holds strategy_config/pipeline_config/moe_config and STRATEGY_MAP moved to _STRATEGY_MAP in config.py; MoEParallelizerConfig now lives in components/distributed/config.py. --------- Signed-off-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> Signed-off-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> Signed-off-by: Adil Asif <adasif@nvidia.com> Signed-off-by: Yuhe Zhang <yuhez@nvidia.com> Signed-off-by: thyways <2484113689@qq.com> Signed-off-by: khazic <khazzz1c@gmail.com> Signed-off-by: Pranav Prashant Thombre <pthombre@nvidia.com> Signed-off-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Signed-off-by: NeMo Bot <nemo-bot@nvidia.com> Signed-off-by: linnan wang <linnanw@nvidia.com> Signed-off-by: HuiyingLi <willwin.lee@gmail.com> Co-authored-by: Adil <47084919+adil-a@users.noreply.github.com> Co-authored-by: Alexandros Koumparoulis <akoumparouli@nvidia.com> Co-authored-by: Alexandros Koumparoulis <153118171+akoumpa@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: Yuhe Zhang <yuhezhang.zju@gmail.com> Co-authored-by: khazzz1c <khazzz1c@gmail.com> Co-authored-by: thyways <2484113689@qq.com> Co-authored-by: Huiying <willwin.lee@gmail.com> Co-authored-by: Pranav Thombre <pthombre@nvidia.com> Co-authored-by: Dong Hyuk Chang <9426164+thomasdhc@users.noreply.github.com> Co-authored-by: NeMo Bot <nemo-bot@nvidia.com> Co-authored-by: jgerh <163925524+jgerh@users.noreply.github.com> Co-authored-by: linnan wang <linnanw@nvidia.com>

khazic requested review from HuiyingLi, ZhiyuLi-Nvidia, adil-a, akoumpa, athitten, hemildesai, pthombre and zyzhou5 as code owners June 7, 2026 08:25

github-actions Bot added the community-request label Jun 7, 2026

copy-pr-bot Bot had a problem deploying to test June 7, 2026 08:28 Error

copy-pr-bot Bot temporarily deployed to nemo-ci June 7, 2026 08:28 Inactive

copy-pr-bot Bot had a problem deploying to nemo-ci June 7, 2026 08:28 Error

copy-pr-bot Bot temporarily deployed to nemo-ci June 7, 2026 08:28 Inactive

copy-pr-bot Bot temporarily deployed to public June 7, 2026 08:29 Inactive

copy-pr-bot Bot temporarily deployed to public June 7, 2026 08:31 Inactive

copy-pr-bot Bot temporarily deployed to nemo-ci June 7, 2026 08:33 Inactive

copy-pr-bot Bot had a problem deploying to nemo-ci June 7, 2026 08:33 Error

copy-pr-bot Bot temporarily deployed to nemo-ci June 7, 2026 08:33 Inactive

copy-pr-bot Bot temporarily deployed to public June 7, 2026 08:39 Inactive

khazic force-pushed the khazic/feat/eagle3-sequence-packing branch from 14a8e7f to 160078c Compare June 7, 2026 08:43

copy-pr-bot Bot temporarily deployed to nemo-ci June 7, 2026 08:52 Inactive

copy-pr-bot Bot temporarily deployed to public June 7, 2026 08:59 Inactive

Merge branch 'main' into khazic/feat/eagle3-sequence-packing

6ce90d1

copy-pr-bot Bot temporarily deployed to nemo-ci June 8, 2026 03:48 Inactive

copy-pr-bot Bot temporarily deployed to test June 8, 2026 03:48 Inactive

copy-pr-bot Bot temporarily deployed to public June 8, 2026 03:48 Inactive

copy-pr-bot Bot temporarily deployed to public June 8, 2026 03:52 Inactive

copy-pr-bot Bot temporarily deployed to nemo-ci June 8, 2026 03:53 Inactive

copy-pr-bot Bot temporarily deployed to public June 8, 2026 03:53 Inactive

copy-pr-bot Bot temporarily deployed to public June 8, 2026 04:01 Inactive

svcnvidia-nemo-ci added the waiting-on-customer Waiting on the original author to respond label Jun 8, 2026

HuiyingLi approved these changes Jun 8, 2026

View reviewed changes

HuiyingLi enabled auto-merge (squash) June 8, 2026 15:50

HuiyingLi merged commit f69851e into NVIDIA-NeMo:main Jun 8, 2026
66 checks passed

svcnvidia-nemo-ci removed the waiting-on-customer Waiting on the original author to respond label Jun 9, 2026

khazic mentioned this pull request Jul 7, 2026

[Tracking] Speculative decoding: implemented features and remaining gaps #2958

Open

24 tasks

Navigation Menu

Uh oh!

feat(speculative): add EAGLE-3 sequence packing and reasoning-mode control - #2444

feat(speculative): add EAGLE-3 sequence packing and reasoning-mode control#2444
HuiyingLi merged 3 commits into
NVIDIA-NeMo:mainfrom
khazic:khazic/feat/eagle3-sequence-packing

khazic commented Jun 7, 2026

Uh oh!

copy-pr-bot Bot commented Jun 7, 2026

Uh oh!

HuiyingLi commented Jun 7, 2026

Uh oh!

HuiyingLi commented Jun 7, 2026

Uh oh!

HuiyingLi commented Jun 8, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

khazic commented Jun 7, 2026

What does this PR do ?

Changelog

Before your PR is "Ready for review"

Additional Information

Uh oh!

copy-pr-bot Bot commented Jun 7, 2026

Uh oh!

HuiyingLi commented Jun 7, 2026

Uh oh!

HuiyingLi commented Jun 7, 2026

Uh oh!

HuiyingLi commented Jun 8, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants