chankhavu/c2.eagle3-test

Eagle3 Long-Context Draft Head for Nemotron-Cascade-2-30B-A3B (Sliding-Window 4k)

This is an Eagle3 speculative-decoding draft head trained against nvidia/Nemotron-Cascade-2-30B-A3B as the verifier. To our knowledge it is the first Eagle3 head trained against a hybrid Mamba-Transformer MoE verifier, and the first to explore sliding-window attention on the draft as a long-context optimization.

TL;DR

• Verifier: nvidia/Nemotron-Cascade-2-30B-A3B (30B-param hybrid Mamba-Transformer MoE, 52 layers = 23 Mamba + 23 MLP/MoE + 6 GQA attention, per the verifier's hybrid_override_pattern config).
• Draft architecture: 1-layer Llama transformer block with hidden_size=2688 (matches verifier residual stream), intermediate_size=8064 (~3x hidden), head_dim=128, num_attention_heads=32, num_key_value_heads=2 (GQA), sliding_window=4096 (band causal), max_position_embeddings=262144 (matches verifier so vLLM does not clamp serving max_model_len), vocab_size=131072, draft_vocab_size=32000.
• Aux hidden state layers captured from verifier: layers 2 / 26 / 48 (~4% / 51% / 94% depth). See "Layer indexing" section below for the exact mapping and the engine-portability caveat. Layout follows NVIDIA's gpt-oss-120b long-context Eagle3 reference.

Layer indexing -- which verifier layers feed the Eagle3 draft

Eagle3 draft heads consume hidden states from THREE specific layers of the verifier (early / middle / late) and learn to map them into the draft's prediction. The choice of which 3 layers matters: too early and the draft has no semantics; too late and it's just predicting from the verifier's own next-token distribution.

This draft is trained against verifier layers 2 / 26 / 48 out of the verifier's 52 hybrid layers. The indexing convention is critical and engine-dependent.

Verifier layer pattern (canonical, from HF config.json)

NemotronH publishes its hybrid layer arrangement as hybrid_override_pattern in the verifier's config.json. For nvidia/Nemotron-Cascade-2-30B-A3B:

hybrid_override_pattern = "MEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEM*EMEMEMEM*EMEMEMEME"

where each character corresponds to one transformer block in the 52-layer stack (0-indexed):

• M = Mamba (state-space) layer
• E = MLP / MoE feed-forward layer
• * = GQA attention layer

Counts: 23 M + 23 E + 6 * = 52.

What layers 2 / 26 / 48 actually are

Index counting all 52 transformer blocks starting from 0 (matches HF transformers model.backbone.layers[i] ordering):

Index	Char	Type	Depth	Notes
2	M	Mamba	~4%	Early Mamba (3rd block overall, 2nd Mamba block)
26	*	Attention	~51%	Middle attention -- the 4th of 6 attention layers, central position
48	M	Mamba	~94%	Late Mamba (3rd-from-last block)

The 6 attention layers are at canonical indices [5, 12, 19, 26, 33, 42]. Layer 26 is the central one.

Engine-portability warning

Different inference engines may number layers differently because of how they fuse, group, or skip blocks. Before serving this draft on a new engine, verify that the engine's "layer i" corresponds to the same character in the hybrid pattern as HF transformers does.

• vLLM (0.19+): uses the full 52-layer ModuleList in its NemotronH model implementation, 0-indexed contiguous, matches HF transformers. The target_hidden_state_indices: [2, 26, 48] in this draft's config.json is read directly by vLLM's Eagle3 speculative decoder and passed to the verifier as raw indices into model.backbone.layers. Tested working on vLLM 0.19 with the in-flight L=32k iteration.

• SGLang: SGLang's NemotronH implementation also uses 0-indexed contiguous layer counting that matches HF, but as of late 2025 it does NOT yet support sliding-window attention on Eagle3 drafts (this draft has sliding_window: 4096 in its config). Once SGLang ships sw support, the layer indices should map cleanly. Until then, prefer vLLM for serving this draft.

• HF transformers (raw, no engine): the verifier exposes its layers as model.backbone.layers[0..51], where layers[i] is the i-th character of hybrid_override_pattern. This is the canonical reference. The training pipeline in this fork (SpecForge with the NemotronH backend patch) reads from these indices via the aux_hidden_state_layers config field on the draft.

• TensorRT-LLM / others: NOT tested. If you port to one of these, cross-check by running a single forward pass on a fixed input and comparing the hidden state dump at indices 2/26/48 against vLLM's dump for the same input. The hidden states should be bit-equivalent (or near-equivalent up to dtype) at the same indices if the numbering convention matches.

Why these specific indices

These three positions follow NVIDIA's gpt-oss-120b long-context Eagle3 recipe (early/mid/late triad). The choice of 2 / 26 / 48 as opposed to e.g. 0 / 26 / 51 was made to:

• Avoid layer 0 (the embedding-adjacent layer), where hidden states are still mostly raw token embeddings without much contextualization.
• Skip the very last layer (51), where hidden states are essentially the verifier's own pre-lm_head representation -- the draft would learn a near-identity from there to the verifier's output and not generalize.
• Pick a Mamba+Attention+Mamba triad rather than three of the same type, so the draft sees a mix of state-space (long-range, Mamba) and attention (short-range, GQA) representations of the same input.
• Trainable parameters: ~196 M (excludes the frozen embedding layer loaded from the verifier).
• Checkpoint: epoch_2_step_12000 (epoch 2, step 12000).

Training schedule: 2-stage (SFT bootstrap + on-policy specialization)

This draft head is trained in two stages to first build a general-language foundation and then specialize to the on-policy reasoning trace distribution that the verifier will actually emit at inference time.

Stage	Source	Rows	LR	Epochs	Bucketing	Notes
Stage 1 (SFT bootstrap)	cascade2_sft_train.jsonl	20,000	1e-4	1	on (length-bucketed sampler, stage 1 only)	Broad SFT distribution; long-tail length variance, so length-bucketed sampling kills the straggler-rank bottleneck on DP=2
Stage 2 (on-policy fine-tune)	c2_traces_train.jsonl	9,658	5e-5	2-3	off	Loaded from stage 1 checkpoint via --ckpt-dir (weight-only, fresh optimizer). Bucketing intentionally disabled to maximize gradient diversity per step on the narrower on-policy distribution we're specializing to.

The CoT trace set (c2_traces_cot_train.jsonl, 0.9k rows) is excluded from this iteration because it has significantly higher mean tokens-per-row (50k) than the plain traces (21k), which would slow stage 2 wall-clock for marginal benefit on the first baseline. It will be folded back in for a follow-up stage 2 fine-tune if the deployed draft underperforms on CoT-heavy reasoning.

Key training details

Training framework	SpecForge fork at https://github.com/hav4ik/SpecForge-Nemotron3, branch nemotron-cascade-2-experiments
Git commit	f517199ed2db
Training date	2026-04-09 UTC
Max sequence length	32768 tokens (this iteration is the L=32768 fast-iteration baseline; an L=65536 follow-up is planned once this baseline ships -- see "Iteration plan" section below)
Truncation loss at this max_length	Stage 1: ~26% of tokens lost to L=32768 truncation (avg tok/row 18904 → 13898). Stage 2: ~15% of tokens lost (avg 21468 → 18179). See "Training data accounting" section below for the per-stage table.
TTT length	6 (Eagle3 test-time-training unroll depth)
Sliding window	4096 tokens on the draft attention
Optimizer	BF16Optimizer, lr=5e-05, warmup_ratio=0.02
Epochs (this stage)	3
Per-rank batch size	1 (no sequence packing)
Tensor parallel	TP=1, DP=2
Global effective batch per step	2 conversations

Training data accounting (the full token-count picture)

Three different but equally valid token counts exist for this dataset, and they don't agree because they measure different things. Documented here to prevent future confusion:

Metric	Stage 1 (cascade2_sft_train)	Stage 2 (c2_traces_train)	Source
Rows / conversations	20,000	9,658	dataset card + jsonl line count
Untruncated total tokens (1)	378,073,468	207,336,404	dataset card (apply_chat_template(messages, tokenize=True))
Truncated total tokens at L=32768 (2)	277,963,707	175,570,455	this fork's tokenizer pipeline w/ max_length=32768
Truncation loss vs (1)	100 M / 26.5%	32 M / 15.3%	(1) - (2)
Loss-masked tokens (assistant turns only, what loss is computed on) (3)	132,941,808	158,793,164	this fork, loss_mask == 1 positions in the tokenized sequence
Loss-mask fraction of (2)	47.8%	90.4%	(3) / (2)
Avg untruncated tok/row	18,904	21,468	dataset card
Avg truncated tok/row	13,898	18,179	(2) / rows
Avg loss-masked tok/row	6,647	16,442	(3) / rows
jsonl size on disk	1521 MiB	660 MiB	filesystem

The three counts measure:

1. Untruncated total -- every token in the conversation including system prompt + user turns + assistant turns + special tokens, with no max-length cap. This is what the dataset card publishes.
2. Truncated total -- same thing but capped at the training max_length. Any sample longer than the cap loses its tail. Stage 1 loses ~26% of total tokens at L=32768 because most SFT conversations are >18k tokens and many exceed 32k. Stage 2 loses ~15%. This is the main reason an L=65536 follow-up is planned.
3. Loss-masked -- the subset of (2) where loss_mask == 1, i.e. only the assistant-turn positions. These are the only positions Eagle3's distillation loss is computed on. Stage 2's 90% loss-mask fraction reflects that reasoning traces are mostly the model's output (short user prompt, very long assistant trace). Stage 1's 48% reflects more balanced SFT dialogue.

Eagle3 vocabulary pruning (the union d2t / t2d mapping)

Eagle3's draft head predicts logits over a smaller draft vocabulary than the verifier's full vocab to keep the lm_head shape manageable. Concretely: target_vocab_size=131072 → draft_vocab_size=32000. The mapping (d2t and t2d buffers on the draft model) is the top-32000 most frequent target tokens in the training set, computed from loss-masked positions.

Multi-stage gotcha (fixed in this fork): the standard SpecForge pipeline auto-generates the d2t / t2d from whatever training set the current run is processing, and unconditionally calls draft_model.load_vocab_mapping(...) AFTER loading the checkpoint passed via --ckpt-dir. For 2-stage training this would mean stage 2 OVERWRITES the d2t / t2d buffers loaded from the stage 1 checkpoint with a NEW mapping derived from stage 2's train set. The lm_head weights loaded from stage 1 would still be aligned to stage 1's mapping → silent index permutation mismatch in every gradient step of stage 2.

Fix: a union vocab mapping built from the combined token frequencies of stage 1 + stage 2, saved once before training, and passed to BOTH stage launchers via --vocab-mapping-path. This way the lm_head is index-aligned end to end.

Per-stage coverage with the union top-32000:

Stage	Total loss-masked tokens	Unique token ids	Coverage by union top-32k	Tokens lost (out of vocab)	Unique-in-vocab
Stage 1	132,941,808	81,439	99.5328%	621,131 (0.47%)	31,978 / 81,439 (39.27%)
Stage 2	158,793,164	36,510	99.9621%	60,180 (0.04%)	26,017 / 36,510 (71.26%)
Union	291,734,972	83,209	99.77%	671,311 (0.23%)	32,000 / 83,209 (38.46%)

Stage 2 has better coverage than stage 1 even though the union top-K was built jointly -- because stage 2's reasoning-trace distribution is much narrower (~half the unique tokens of stage 1 despite ~20% more total tokens). Stage 2 contributes 1,770 unique tokens not present in stage 1's loss-masked positions (mostly latex / reasoning markers); these would have been silently dropped from the draft vocab if the mapping had been built from stage 1 only.

Both stages comfortably exceed 99% frequency coverage with draft_vocab_size=32000, so the chosen draft vocab is big enough -- no need to grow it.

Iteration plan: this is the L=32768 fast-iteration baseline

This checkpoint is the first Eagle3 head trained against Nemotron-Cascade-2 in this fork's iteration sequence:

Iteration	max_length	Status	Goal
L=32768 (this)	32768	in flight / shipped	Fast-iteration baseline. Get a working draft with all the engineering knobs validated end-to-end. Pays a ~26% truncation loss on stage 1 and ~15% on stage 2.
L=65536 (next)	65536	planned	Long-context follow-up. Re-train both stages at the higher cap to recover the truncated tail. Will need --draft-mlp-grad-checkpoint re-enabled to fit on a 96 GB GPU per rank.

The per-length eval data (16k / 32k / 64k validation sets, see below) lets us measure how well the L=32k draft generalizes to longer contexts without training at long context. If the L=32k → L=64k generalization gap on the eval set is small, the L=65k follow-up may not be worth re-training. If the gap is large, the follow-up is justified.

Note on batch size + length-bucketed sampling

This draft is trained with per-rank batch_size=1 and no sequence packing. With dp_size=2 data-parallel ranks the global effective batch is 2 conversations per optimizer step. This is the standard configuration for long-context Eagle3 training (NVIDIA's gpt-oss-120b long-context Eagle3 uses the same per-rank batch=1 layout) -- packing into fixed-token-budget batches gives diminishing returns at long context because each conversation already saturates the GPU activation budget.

Per-step token counts vary by ~20x because conversation lengths in the training distribution range from ~30 to ~30000+ tokens. Without length-aware sampling, every optimizer step is bottlenecked by the slowest rank: a (2k, 32k) DP-pair wastes half the GPU while rank 1 sits idle waiting for rank 0.

Stage 1 fixes this with length-bucketed sampling (--with-data-bucketing, see specforge/data/utils.py:LengthBucketDistributedSampler): the sampler sorts the dataset by per-sample length, partitions the sorted index into contiguous "global batches" of size num_replicas * batch_size, shuffles only the order of those global batches per epoch with a deterministic seed, and assigns each rank its slice of every global batch. Effect: all ranks within a single optimizer step see samples of similar length, eliminating the straggler-rank waste.

Bias mitigation: pure length-sorted iteration would bias gradient updates by difficulty across an epoch (long samples first or last). The per-epoch global-batch-order shuffle mitigates this; the intra-batch order is intentionally NOT shuffled so each step still sees bucketed lengths. Same tradeoff as HF Trainer's group_by_length=True.

Stage 2 leaves bucketing OFF intentionally. The on-policy reasoning trace distribution is narrow enough that gradient diversity per step matters more than throughput, and the per-step difficulty drift would work against the "specialize precisely" goal of stage 2.

| Wandb run | https://wandb.ai/hav4ik/nemotron-cascade-2-eagle3/runs/yz5vd3qw |

⚠️ CRITICAL: Mamba SSM state precision

This draft was trained with the Mamba SSM scan boundary state in float32, matching vLLM 0.19's default behavior for NemotronH (--mamba_ssm_cache_dtype float32).

The Nemotron team has explicitly confirmed (HF discussion) that downcasting the Mamba SSM state to bf16 causes a ~10% absolute regression on AIME-class math benchmarks (88.3% vs 99.17% on AIME 2025 was the SGLang-vs-vLLM gap they measured from this single issue alone).

Both training and inference must use fp32 SSM state to avoid this precision regression. Concretely:

• Training: SpecForge fork includes a monkey-patch (specforge/_mamba_fp32_patch.py) that forces upstream mamba_ssm's _state_passing_fwd to use out_dtype=torch.float32. The patch is applied at the very top of scripts/train_eagle3.py before any other module-level imports. It is bit-equivalent to vLLM 0.19's NemotronH default behavior.
• Inference (vLLM 0.19+): works automatically for NemotronH-Cascade-2 because the verifier's config.json ships with mamba_ssm_cache_dtype: "float32", which triggers vLLM's NemotronHForCausalLMConfig.verify_and_update_config to set the SSM cache to fp32. No explicit flag needed when serving NemotronH-Cascade-2 with vLLM 0.19+.
• Inference (SGLang): pass --mamba-ssm-dtype float32 (the SGLang flag is not on by default, per the Nemotron team's recommendation).
• Inference (other frameworks / older vLLM): explicitly pass --mamba_ssm_cache_dtype float32 (vLLM) or equivalent.

For the full audit story (call chain, the exact line numbers in upstream mamba_ssm 2.3.1 where the bf16 downcast happens, vLLM's 4-layer override chain, etc.), see the HANDOFF.md "Mamba SSM precision -- the full intricacies" section.

Inference: serving with vLLM

vllm serve nvidia/Nemotron-Cascade-2-30B-A3B \
    --speculative-config '{
        "model": "<path-to-this-checkpoint-or-hf-repo>",
        "method": "eagle3",
        "num_speculative_tokens": 5
    }' \
    --trust-remote-code \
    --max-model-len 262144 \
    --tensor-parallel-size 2

vLLM 0.19+ will automatically:

• Load the verifier with fp32 Mamba SSM state (auto-read from the verifier's config.json -- no explicit flag needed)
• Apply sliding-window attention to the draft using the sliding_window: 4096 and layer_types: ["sliding_attention"] fields in this checkpoint's config.json. The Eagle3-aware target_layer_count offset is auto-computed from the verifier (no need to set it manually).
• Bound the draft KV cache to the 4096-token window via vLLM's SlidingWindowSpec in the hybrid KV cache manager.

Inference: serving with SGLang

SGLang does not yet support sliding-window attention on Eagle3 drafts (as of late 2025 -- check the SGLang changelog for status). When SGLang adds support, the equivalent invocation will be:

python -m sglang.launch_server \
    --model nvidia/Nemotron-Cascade-2-30B-A3B \
    --speculative-algorithm EAGLE3 \
    --speculative-draft-model-path <path-to-this-checkpoint-or-hf-repo> \
    --speculative-num-steps 6 \
    --speculative-eagle-topk 4 \
    --speculative-num-draft-tokens 16 \
    --mamba-ssm-dtype float32 \
    --trust-remote-code \
    --tp 2

Until SGLang ships sliding-window support for Eagle3 drafts, prefer serving this draft via vLLM 0.19+.

Sliding-window attention -- the experimental hypothesis

The draft is a 1-layer Llama transformer; without windowing its self-attention is O(L^2) per query position. The verifier (Nemotron-Cascade-2) only has 6 attention layers and 31 Mamba layers, so the verifier's compute scales gracefully with context length, but the Eagle3 draft becomes the long-context bottleneck.

Hypothesis: most reasoning-trace tokens are local-pattern continuations -- the draft only needs to predict locally-coherent next tokens; the verifier rejects any draft proposal that diverges globally. So the draft should be able to do well with a 4096-token sliding window, collapsing its attention cost from O(L^2) to O(L*window) and freeing us to push training context length up.

This checkpoint is the first published Eagle3 head built around this hypothesis. Acceptance rates / downstream speedup numbers (when available) will be added to a follow-up section here.

Memory engineering required to train at long context

Training a 1-layer Llama-style draft against a 30B-param frozen Mamba-Transformer verifier at long context (L=32k or L=65k) with ttt_length=6 does not fit on a 96 GB GPU per rank without significant engineering. The SpecForge fork that produced this checkpoint adds the following training-time optimizations:

1. Chunked MLP forward (--draft-mlp-chunk-size 4096, active for this checkpoint): Liger-style per-chunk MLP forward over the seq dim, drops the per-step transient peak from ~4 GiB to ~800 MiB per MLP forward. Cheap defensive memory saving with negligible compute overhead.
2. Chunked fused linear + soft-target CE (--fused-linear-loss --fused-linear-loss-chunk-size 4096, active for this checkpoint): chunks the lm_head + KL-distillation loss over the seq dim with per-chunk grad checkpointing, never materializes the full [B, T, V] logits tensor. Saves ~14 GiB at L=32k and ~28 GiB at L=65k. Mathematically equivalent to the unchunked path (numerical equivalence verified in the SpecForge fork's specforge/core/loss.py __main__ block).
3. Sliding-window attention on the draft (sliding_window=4096, active for this checkpoint): collapses draft self-attention from O(L²) to O(L * window). See "Sliding-window attention" section.
4. Mamba SSM fp32 monkey-patch (always on): see Mamba SSM section.
5. Patch to free verifier full-vocab logits before draft TTT unrolling so the ~16 GiB [L, V] tensor doesn't stay alive across the backward pass.
6. In-place padding for the per-step shift operations on big logit tensors.

Disabled at L=32k, re-enabled at L=65k:

7. Grad-checkpoint on the draft MLP (--draft-mlp-grad-checkpoint, OFF for this L=32k checkpoint, will be ON for the L=65k follow-up): recomputes gate_proj/silu/up_proj/(gate*up) intermediates during backward, drops them from the saved-for-backward set across TTT unrolls. Required at L=65k to fit on 96 GB but costs ~25-30% on the draft forward, so we leave it off at L=32k where we have memory headroom and want maximum step throughput.

See the experiments/nemotron-cascade-2/README.md for the complete memory math reference and the per-commit engineering tour.

License and attribution

• Verifier model (nvidia/Nemotron-Cascade-2-30B-A3B): refer to the verifier's HF page for its license. This draft head is derivative in the sense that it was trained against the verifier's outputs.
• This draft head: Apache 2.0 (matching the SpecForge license).
• Training framework: SpecForge fork at https://github.com/hav4ik/SpecForge-Nemotron3, derived from https://github.com/sgl-project/SpecForge with additional engineering patches for hybrid Mamba-Transformer verifiers (see branch nemotron-cascade-2-experiments).