Train an EAGLE Drafter for Speculative Decoding — End-to-End Guide

A step-by-step guide for training an EAGLE speculative decoding drafter to accelerate LLM inference using NeMo AutoModel.

What is EAGLE Speculative Decoding?

Large language models generate text one token at a time — each token requires a full forward pass through the entire model. Speculative decoding speeds this up by pairing the large target model with a small, fast drafter model. The drafter guesses multiple tokens ahead; the target model then verifies them all in a single forward pass, accepting correct guesses and rejecting wrong ones. The output is mathematically identical to running the target model alone, but 2-3x faster.

EAGLE (Extrapolation Algorithm for Greater Language-model Efficiency) is a family of speculative decoding methods. NeMo AutoModel supports three variants:

The Task

We train an EAGLE-3 drafter for Llama 3.1 8B Instruct on the PerfectBlend dataset — a chat corpus whose assistant turns were generated by the same Llama 3.1 8B model, ensuring distribution alignment between training data and target.

After training, we serve the target + drafter together via SGLang for accelerated inference.

Guide Overview

Hardware Requirements

Step 0 — Environment Setup

This guide runs inside the NeMo AutoModel Docker container:

$
docker run -it --rm --gpus all --ipc=host --network host \
>
    -v $(pwd):/workspace \
>
    nvcr.io/nvidia/nemo-automodel:26.06.00

$
huggingface-cli login

$
cd /opt/Automodel

For SGLang serving (Step 5), install it in the same environment:

$
uv pip install "sglang>=0.5.9"

Step 1 — Understand EAGLE Architecture

How EAGLE-3 Works

EAGLE-3 pairs a frozen target LLM with a small trainable drafter. During training, the drafter learns to predict what the target model would produce next, using a technique called test-time training (TTT) unroll:

Target (frozen)          Drafter (trainable)
┌──────────────┐         ┌──────────────┐
│ Llama 3.1 8B │ ──────> │  1-layer     │
│              │ hidden  │  transformer │
│ Full model   │ states  │  + fc fusion │
│ (frozen)     │         │  + lm_head   │
└──────────────┘         └──────────────┘
                               │
                         predict next token
                         + hidden states

Key components:

• Target model: The full LLM (e.g., Llama 3.1 8B), completely frozen. Provides hidden states from selected intermediate layers as auxiliary inputs to the drafter.
• Draft model: A shallow transformer (a single fused decoder layer) with:
  • A fusion layer (fc) that combines auxiliary hidden states from 3 target layers
  • Its own attention layers, MLP, and layer norm
  • A smaller output vocabulary (e.g., 8192 or 32000 tokens instead of 128k) to reduce compute
• TTT unroll: The drafter runs the configured number of autoregressive steps during training; the maintained Llama examples use 4. Loss weights decay as 0.8^i, teaching the drafter to make multi-step predictions — exactly what speculative decoding needs.

Supported Target Architectures

The target is resolved from its HF config.architectures string through the registry in components/speculative/eagle/registry.py. Most dense and MoE decoders reuse the config-driven Llama-style draft (a one-line registry entry); architectures whose RoPE or attention layout the shared draft cannot represent get a thin dedicated draft class.

Gemma4. Gemma4 ships as a multimodal Gemma4ForConditionalGeneration whose text decoder config is nested under config.text_config; the draft is built from that inner config automatically. Set target_force_hf: true so the stock HuggingFace text-capable model is loaded (the draft hooks the decoder under model.language_model.layers). The draft consumes only post-block hidden states, so it is the same Llama-style dense draft used elsewhere, with two Gemma quirks reconciled: the hidden_activation (GeGLU) key and Gemma4’s nested per-attention-type rope_parameters, which is flattened to a standard full-rotary Llama RoPE on the global (full-attention) base. The saved checkpoint keeps the canonical architectures: ["LlamaEagle3DraftModel"] layout, so SGLang and vLLM serve it with their existing EAGLE-3 Llama head. See examples/speculative/eagle3/gemma4_e2b_eagle3.yaml.

EAGLE-3.1 Drafter Toggles

The same train_eagle3 recipe supports the EAGLE-3.1 drafter variant via two optional flags in recipe_args. Both default to false, so existing EAGLE-3 configs and checkpoints behave identically. Setting them applies the EAGLE-3.1 architectural changes from vllm-project/vllm#42764 to the Llama-style draft. The MLA-backbone community release lightseekorg/kimi-k2.6-eagle3.1-mla is a separate architecture (Eagle3DeepseekV2ForCausalLM) and is not produced by this recipe.

Together they remove the “attention drift” pattern (loss of focus on sink tokens at deeper speculation depths) reported by the EAGLE-3.1 paper and let the drafter behave more like a recurrently applied module than a stack of extra layers bolted onto the target.

1
recipe_args:
2
  # ... standard EAGLE-3 fields ...
3
  fc_norm: true
4
  norm_output: true

How EAGLE-1 Differs

EAGLE-1 is simpler: it uses a single transformer layer, predicts the full vocabulary, and trains with a combined loss of SmoothL1 (Huber) on hidden states (hidden_loss_weight) and cross-entropy on tokens (token_loss_weight). No TTT unroll, no vocabulary mapping.

Step 2 — Prepare the Training Dataset

Data format

EAGLE training accepts chat data in the OpenAI messages format—either JSONL files or HuggingFace datasets with a messages column—and automatically converts supported ShareGPT-style conversations rows to that format:

1
{"messages": [
2
  {"role": "system", "content": "You are a helpful assistant."},
3
  {"role": "user", "content": "What is the capital of France?"},
4
  {"role": "assistant", "content": "The capital of France is Paris."}
5
]}

Option A: Use a pre-regenerated dataset (recommended)

For best results, the assistant turns in your training data should come from the same model you’ll use as the target at inference time. The PerfectBlend dataset already has answers regenerated by Llama 3.1 8B Instruct:

$
python -c "
>
from datasets import load_dataset
>
ds = load_dataset(
>
    'frankleeeee/PerfectBlend-Regenerated-Llama-3.1-8B-Instruct',
>
    split='train[:5]'
>
)
>
print(f'Columns: {ds.column_names}')
>
print(f'Sample conversation:')
>
for msg in ds[0]['conversations'][:3]:
>
    role = msg['role']
>
    text = msg['content'][:80]
>
    print(f'  [{role}] {text}...')
>
"

Expected output:

Columns: ['conversations']
Sample conversation:
  [system] You are a helpful assistant....
  [user] What are the main differences between Python 2 and Python 3?...
  [assistant] Here are the key differences between Python 2 and Python 3:
1. **P...

Option B: Regenerate answers from your target model

If you have a chat dataset whose answers were generated by a different model, you can regenerate them using your target. This aligns the training data distribution with the model the drafter will actually assist at inference time.

Step B.1 — Start the target server (in one shell):

$
python -m sglang.launch_server \
>
    --model-path meta-llama/Llama-3.1-8B-Instruct \
>
    --port 30000

Wait for Uvicorn running on http://0.0.0.0:30000 before proceeding.

Step B.2 — Regenerate (in another shell):

$
python -m nemo_automodel.components.speculative.regenerate \
>
    --input-data Aeala/ShareGPT_Vicuna_unfiltered \
>
    --output-dir ./regenerated/sharegpt_llama31_8b \
>
    --target-server http://localhost:30000/v1 \
>
    --model meta-llama/Llama-3.1-8B-Instruct \
>
    --concurrency 64 \
>
    --shard-size 1000

For each sample, the script:

1. Loads the conversation from the input dataset
2. Drops the trailing assistant turn, keeping the user prompt context
3. Calls the target server to generate a new assistant response
4. Saves the rebuilt conversation to parquet shards

The output directory contains parquet files with a messages column — ready for EAGLE training. The script is resumable: rerun with --resume to skip completed shards.

Step 3 — Configure and Launch EAGLE-3 Training

YAML config

Save the following as eagle3_llama8b.yaml:

1
recipe: TrainEagle3Recipe
2
3
dist_env:
4
  backend: nccl
5
  timeout_minutes: 60
6
7
recipe_args:
8
  target_model_name_or_path: meta-llama/Llama-3.1-8B-Instruct
9
10
  # Point to your training data (HF dataset id, local parquet dir, or JSONL)
11
  train_data_path: frankleeeee/PerfectBlend-Regenerated-Llama-3.1-8B-Instruct
12
  val_data_path: null
13
14
  # Slice to 200k samples for a ~2h training run
15
  train_split: "train[:200000]"
16
  val_split: null
17
18
  output_dir: ./outputs/eagle3_llama8b
19
  seq_length: 2048
20
  micro_batch_size: 1
21
  grad_accumulation_steps: 4   # effective batch = 8 GPUs * 1 * 4 = 32
22
  num_workers: 4
23
  num_epochs: 1
24
25
  # EAGLE-3 specific
26
  ttt_steps: 4                 # TTT unroll depth (higher = better but slower)
27
  draft_vocab_size: 32000      # smaller vocab = faster drafter
28
29
  # The drafter copies the target's embedding table at init; this flag
30
  # freezes those copied weights so only the draft transformer layers
31
  # and lm_head are trained.
32
  freeze_embeddings: true
33
  shuffle_seed: 42
34
  log_every_steps: 20
35
  max_grad_norm: 1.0
36
37
optimizer:
38
  lr: 2.0e-4
39
  betas: [0.9, 0.95]
40
  weight_decay: 0.0
41
  warmup_ratio: 0.05           # 5% warmup
42
  min_lr_ratio: 0.1
43
44
checkpoint:
45
  enabled: true
46
  checkpoint_dir: ./outputs/eagle3_llama8b/checkpoints
47
  # The recipe defaults to safetensors + consolidated; these lines are
48
  # shown explicitly for clarity but can be omitted.
49
  model_save_format: safetensors
50
  save_consolidated: true

Config field reference

Launch training

Multi-GPU (8x A100, production):

$
torchrun --nproc-per-node=8 \
>
    -m nemo_automodel.recipes.llm.train_eagle3 \
>
    -c eagle3_llama8b.yaml

Single-GPU (quick test with Llama 3.2 1B):

For a quick test, use the MVP config with Llama 3.2 1B and a small dataset:

$
python -m nemo_automodel.recipes.llm.train_eagle3 \
>
    -c examples/speculative/eagle3/llama_eagle3_mvp.yaml

Step 4 — Monitor Training and Inspect Checkpoints

What to watch

At each logging interval, the recipe emits epoch, step, train_loss, train_acc, an optional train_tau_sim, and lr. The values below illustrate the log format:

epoch=0 step=20 train_loss=3.452100 train_acc=0.214500 train_tau_sim=0.3812 lr=8.000e-05
epoch=0 step=40 train_loss=2.897300 train_acc=0.267100 train_tau_sim=0.4198 lr=1.600e-04

Checkpoint layout

The shown 200k-sample, 8-GPU config does not enable step or epoch checkpoint cadence, so it saves only the final checkpoint after 6250 optimizer steps:

outputs/eagle3_llama8b/checkpoints/
  epoch_1_step_6250/
    model/                   # HF-style draft artifacts
    eagle_meta.pt            # EAGLE metadata and token mapping
    optim/                   # optimizer and scheduler state
    rng/                     # distributed RNG state
    config.yaml              # training config snapshot
  LATEST -> epoch_1_step_6250

Resume from checkpoint

To create intermediate checkpoints that can recover an interrupted run, first set a cadence in recipe_args, for example ckpt_every_steps: 1000. Then resume from the latest checkpoint:

1
checkpoint:
2
  restore_from: LATEST

Or point to a specific checkpoint:

1
checkpoint:
2
  restore_from: epoch_0_step_3000

Step 5 — Serve with SGLang

The serve_sglang helper resolves the current HF-style training checkpoint and launches the server in one command. It also converts older bare-weight checkpoints as a compatibility fallback.

Launch the server

$
python -m nemo_automodel.components.speculative.serve_sglang \
>
    --target meta-llama/Llama-3.1-8B-Instruct \
>
    --draft ./outputs/eagle3_llama8b/checkpoints/LATEST \
>
    --algorithm EAGLE3 \
>
    --num-steps 3 \
>
    --topk 1 \
>
    --num-draft-tokens 4 \
>
    --port 30000

For a current checkpoint, the helper:

1. Resolves the checkpoint’s HF-style model/ directory directly
2. Rewrites the architecture name for SGLang compatibility when needed (LlamaEagle3DraftModel -> LlamaForCausalLMEagle3)
3. Uses an existing model/speculative_token_map.pt, or generates it from the sibling eagle_meta.pt for EAGLE-3
4. Launches SGLang with the correct speculative decoding flags

For a legacy checkpoint containing draft_model.pt and config.json, the helper exports the weights and rewritten config into model/. For EAGLE-3, it uses the legacy eagle3_meta.pt when present to generate model/speculative_token_map.pt.

Serving parameters

Pass extra SGLang flags after --:

$
python -m nemo_automodel.components.speculative.serve_sglang \
>
    --target meta-llama/Llama-3.1-8B-Instruct \
>
    --draft ./outputs/eagle3_llama8b/checkpoints/LATEST \
>
    --algorithm EAGLE3 \
>
    -- --enable-torch-compile --schedule-conservativeness 1.2

Smoke-test the server

Once you see Uvicorn running on http://0.0.0.0:30000, test it:

$
curl http://localhost:30000/generate \
>
    -H "Content-Type: application/json" \
>
    -d '{
>
      "text": "Hello, my name is",
>
      "sampling_params": {"max_new_tokens": 64}
>
    }'

Expected output:

1
{
2
  "text": "Hello, my name is Sarah and I am a 25-year-old software engineer...",
3
  "meta_info": {
4
    "prompt_tokens": 6,
5
    "completion_tokens": 64,
6
    "accept_length_per_step": 3.2
7
  }
8
}

The accept_length_per_step metric shows how many tokens the target model accepts per speculative step on average. Higher is better — a value of 3.0+ indicates the drafter is accurately predicting the target’s behavior.

OpenAI-compatible endpoint

SGLang also exposes an OpenAI-compatible API:

$
curl http://localhost:30000/v1/chat/completions \
>
    -H "Content-Type: application/json" \
>
    -d '{
>
      "model": "meta-llama/Llama-3.1-8B-Instruct",
>
      "messages": [
>
        {"role": "user", "content": "Explain speculative decoding in one paragraph."}
>
      ],
>
      "max_tokens": 256
>
    }'

Expected output:

1
{
2
  "choices": [{
3
    "message": {
4
      "role": "assistant",
5
      "content": "Speculative decoding is a technique for accelerating autoregressive language model inference. It works by using a small, fast \"draft\" model to predict multiple future tokens, which are then verified in parallel by the larger \"target\" model in a single forward pass. Tokens that match the target model's predictions are accepted, while incorrect tokens are rejected and regenerated. Because verification is cheaper than sequential generation (it processes all candidate tokens simultaneously), the overall throughput increases significantly — typically 2-3x — while producing output that is mathematically identical to running the target model alone."
6
    }
7
  }],
8
  "usage": {
9
    "prompt_tokens": 14,
10
    "completion_tokens": 112
11
  }
12
}

Step 6 — (Bonus) Train an EAGLE-1 Drafter

EAGLE-1 is simpler and faster to train, making it a good starting point for experimentation. It uses a single transformer layer and trains with a combined hidden-state SmoothL1 (Huber) + token cross-entropy loss.

YAML config

Save as eagle1_llama8b.yaml:

1
recipe: TrainEagle1Recipe
2
3
dist_env:
4
  backend: nccl
5
  timeout_minutes: 30
6
7
recipe_args:
8
  target_model_name_or_path: meta-llama/Llama-3.1-8B-Instruct
9
  train_data_path: frankleeeee/PerfectBlend-Regenerated-Llama-3.1-8B-Instruct
10
  val_data_path: null
11
  train_split: "train[:200000]"
12
  val_split: null
13
  output_dir: ./outputs/eagle1_llama8b
14
  seq_length: 2048
15
  micro_batch_size: 1
16
  grad_accumulation_steps: 4
17
  num_workers: 4
18
  num_epochs: 1
19
20
  # EAGLE-1 specific
21
  draft_num_hidden_layers: 1    # number of draft transformer layers
22
  hidden_loss_weight: 1.0       # SmoothL1 (Huber) loss on hidden states
23
  token_loss_weight: 0.1        # cross-entropy loss on tokens
24
25
  freeze_embeddings: true
26
  shuffle_seed: 42
27
  log_every_steps: 10
28
  max_grad_norm: 1.0
29
30
optimizer:
31
  lr: 1.0e-4
32
  betas: [0.9, 0.95]
33
  weight_decay: 0.0
34
35
checkpoint:
36
  enabled: true
37
  checkpoint_dir: ./outputs/eagle1_llama8b/checkpoints
38
  # Defaults to safetensors + consolidated; can be omitted.
39
  model_save_format: safetensors
40
  save_consolidated: true

Launch

$
torchrun --nproc-per-node=8 \
>
    -m nemo_automodel.recipes.llm.train_eagle1 \
>
    -c eagle1_llama8b.yaml

Serve

Use --algorithm EAGLE (not EAGLE3) for EAGLE-1/2 drafters:

$
python -m nemo_automodel.components.speculative.serve_sglang \
>
    --target meta-llama/Llama-3.1-8B-Instruct \
>
    --draft ./outputs/eagle1_llama8b/checkpoints/LATEST \
>
    --algorithm EAGLE \
>
    --num-steps 3 --topk 1 --num-draft-tokens 4 \
>
    --port 30000

EAGLE-1 vs EAGLE-3

Example Configs Reference

Troubleshooting