> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/nemo/automodel/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nemo/automodel/_mcp/server. # Train an EAGLE Drafter for Speculative Decoding — End-to-End Guide > Step-by-step guide for training EAGLE-3 and EAGLE-1 speculative decoding drafters with NeMo AutoModel, from data preparation through SGLang serving. **A step-by-step guide for training an EAGLE speculative decoding drafter to accelerate LLM inference using [NeMo AutoModel](https://github.com/NVIDIA-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: | Variant | Recipe | Description | | ----------- | -------------- | ------------------------------------------------------------------------------------------ | | **EAGLE-1** | `train_eagle1` | Lightweight 1-layer draft transformer; learns to predict target hidden states + next token | | **EAGLE-2** | `train_eagle2` | Same architecture as EAGLE-1 (alias recipe) | | **EAGLE-3** | `train_eagle3` | Advanced drafter with test-time training (TTT) unroll and vocabulary mapping; best speed | ## The Task We train an EAGLE-3 drafter for **Llama 3.1 8B Instruct** on the [PerfectBlend](https://huggingface.co/datasets/frankleeeee/PerfectBlend-Regenerated-Llama-3.1-8B-Instruct) 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](https://github.com/sgl-project/sglang) for accelerated inference. ## Guide Overview | Step | Description | | ---------- | ---------------------------------------- | | **Step 0** | Environment setup | | **Step 1** | Understand EAGLE architecture | | **Step 2** | Prepare the training dataset | | **Step 3** | Configure and launch EAGLE-3 training | | **Step 4** | Monitor training and inspect checkpoints | | **Step 5** | Serve with SGLang | | **Step 6** | (Bonus) Train an EAGLE-1 drafter | ## Hardware Requirements | Setup | Target Model | GPUs | Training Time | | ---------------- | --------------------- | ------------- | ------------------------------ | | MVP (quick test) | Llama 3.2 1B | 1x A100 80 GB | \~10 min (1 epoch, 1k samples) | | Production | Llama 3.1 8B Instruct | 8x A100 80 GB | \~2 h (1 epoch, 200k samples) | The target model is loaded in full precision and frozen during training. Only the small drafter model (a few transformer layers) has trainable parameters, so GPU memory is dominated by the target model size. *** ## Step 0 — Environment Setup This guide runs **inside** the NeMo AutoModel Docker container: ```bash docker run -it --rm --gpus all --ipc=host --network host \ -v $(pwd):/workspace \ nvcr.io/nvidia/nemo-automodel:26.06.00 ``` ```bash huggingface-cli login ``` ```bash cd /opt/Automodel ``` For SGLang serving (Step 5), install it in the same environment: ```bash 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**: ```text Target (frozen) Drafter (trainable) ┌──────────────┐ ┌──────────────┐ │ Llama 3.1 8B │ ──────> │ 2-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 (typically 2 layers) 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 multiple autoregressive steps (default 4) during training, with exponentially decaying loss weights (`0.8^i`). This teaches the drafter to make multi-step predictions — exactly what speculative decoding needs. ### 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](https://github.com/vllm-project/vllm/pull/42764) to the Llama-style draft. The MLA-backbone community release [`lightseekorg/kimi-k2.6-eagle3.1-mla`](https://huggingface.co/lightseekorg/kimi-k2.6-eagle3.1-mla) is a separate architecture (`Eagle3DeepseekV2ForCausalLM`) and is **not** produced by this recipe. | Flag | Effect | | ------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `fc_norm` | Apply an independent RMSNorm to each of the `num_aux_hidden_states` (3 by default) auxiliary hidden-state chunks before they are concatenated and projected by `model.fc`. Stored as an `nn.ModuleList` with on-disk keys `model.fc_norm.0.weight`, `model.fc_norm.1.weight`, ... matching vLLM's layout so checkpoints load directly. | | `norm_output` | Route the existing final RMSNorm (`model.norm`) over the per-step hidden state returned by the drafter so the next TTT step (and `lm_head`) consume the post-norm state instead of the raw decoder output. Adds no parameters. | 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. ```yaml recipe_args: # ... standard EAGLE-3 fields ... fc_norm: true 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 MSE 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 expects chat data in the **OpenAI messages format** — either JSONL files or HuggingFace datasets with a `messages` column: ```json {"messages": [ {"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "What is the capital of France?"}, {"role": "assistant", "content": "The capital of France is Paris."} ]} ``` ### 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: ```bash 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... ``` PerfectBlend uses a `conversations` column, but `ChatDataset` expects `messages`. Rename the column before training: ```bash python -c " import pandas as pd from pathlib import Path src = Path('') dst = Path('./data/perfectblend_renamed') dst.mkdir(parents=True, exist_ok=True) for f in sorted(src.glob('train-*.parquet')): df = pd.read_parquet(f) df = df.rename(columns={'conversations': 'messages'}) df.to_parquet(dst / f.name, index=False) print('Done. Point train_data_path at:', dst) " ``` ### 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): ```bash 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): ```bash 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. | Flag | Default | Notes | | ------------------ | ------- | ------------------------------------------------ | | `--concurrency` | 32 | In-flight requests; raise to saturate the server | | `--shard-size` | 1000 | Rows per parquet file | | `--temperature` | 0.0 | Greedy by default (recommended for EAGLE) | | `--max-new-tokens` | 1024 | Cap per-answer length | | `--split` | `train` | Supports HF slice syntax, e.g., `train[:10000]` | *** ## Step 3 — Configure and Launch EAGLE-3 Training ### YAML config Save the following as `eagle3_llama8b.yaml`: ```yaml recipe: TrainEagle3Recipe dist_env: backend: nccl timeout_minutes: 60 recipe_args: target_model_name_or_path: meta-llama/Llama-3.1-8B-Instruct # Point to your training data (HF dataset id, local parquet dir, or JSONL) train_data_path: ./data/perfectblend_renamed val_data_path: null # Slice to 200k samples for a ~2h training run train_split: "train[:200000]" val_split: null output_dir: ./outputs/eagle3_llama8b seq_length: 2048 micro_batch_size: 1 grad_accumulation_steps: 4 # effective batch = 8 GPUs * 1 * 4 = 32 num_workers: 4 num_epochs: 1 # EAGLE-3 specific ttt_steps: 4 # TTT unroll depth (higher = better but slower) draft_vocab_size: 32000 # smaller vocab = faster drafter # The drafter copies the target's embedding table at init; this flag # freezes those copied weights so only the draft transformer layers # and lm_head are trained. freeze_embeddings: true shuffle_seed: 42 log_every_steps: 20 max_grad_norm: 1.0 optimizer: lr: 2.0e-4 betas: [0.9, 0.95] weight_decay: 0.0 warmup_ratio: 0.05 # 5% warmup min_lr_ratio: 0.1 checkpoint: enabled: true checkpoint_dir: ./outputs/eagle3_llama8b/checkpoints # The recipe defaults to safetensors + consolidated; these lines are # shown explicitly for clarity but can be omitted. model_save_format: safetensors save_consolidated: true ``` ### Config field reference | Field | What It Does | | ---------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | `target_model_name_or_path` | HuggingFace model ID for the frozen target LLM | | `train_data_path` | Path to chat data (HF dataset id, parquet dir, or JSONL) | | `train_split` | Optional HF slice syntax to limit data size | | `seq_length` | Context window length (1024 for quick tests, 2048 for production) | | `micro_batch_size` | Per-GPU batch size | | `grad_accumulation_steps` | Gradient accumulation for larger effective batches | | `ttt_steps` | TTT unroll depth; 4 is the default, cost is linear per step | | `draft_vocab_size` | Draft output vocabulary size; smaller = faster inference | | `freeze_embeddings` | Freeze the embedding table copied from the target so only draft layers train (recommended `true`) | | `target_attn_implementation` | Optional attention backend for the frozen target (e.g. `sdpa`); defaults to HF auto-select. Set `sdpa` if the target's FlashAttention path is broken on your build (e.g. the Qwen3 FA2 `s_aux` crash) | | `fc_norm` | EAGLE-3.1: per-chunk independent RMSNorm (`ModuleList`) on auxiliary hidden states before the `fc` projection (default `false`) | | `norm_output` | EAGLE-3.1: feed the post-`model.norm` hidden state into the next TTT step and `lm_head` (default `false`) | | `warmup_ratio` | Fraction of total steps for LR warmup | ### Launch training **Multi-GPU (8x A100, production):** ```bash torchrun --nproc-per-node=8 \ -m nemo_automodel.recipes.llm.train_eagle3 \ 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: ```bash python -m nemo_automodel.recipes.llm.train_eagle3 \ examples/speculative/eagle3/llama_eagle3_mvp.yaml ``` For GPUs with FlashAttention support, add `draft_attn_implementation: flash_attention_2` to `recipe_args` for faster training. See [llama\_eagle3\_mvp\_flash\_attn.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle3/llama_eagle3_mvp_flash_attn.yaml) for a complete example. *** ## Step 4 — Monitor Training and Inspect Checkpoints ### What to watch Training loss should drop steadily. Here is a sample log from the PerfectBlend 200k run on 8x A100: ``` [epoch 0] step 20 | loss 3.4521 | grad_norm 12.84 | lr 8.00e-05 | tokens/s 2841 [epoch 0] step 40 | loss 2.8973 | grad_norm 8.31 | lr 1.60e-04 | tokens/s 3102 [epoch 0] step 100 | loss 2.1245 | grad_norm 5.62 | lr 2.00e-04 | tokens/s 3254 [epoch 0] step 500 | loss 1.5832 | grad_norm 3.17 | lr 1.98e-04 | tokens/s 3198 [epoch 0] step 1000 | loss 1.3401 | grad_norm 2.45 | lr 1.92e-04 | tokens/s 3221 [epoch 0] step 3000 | loss 1.1056 | grad_norm 1.89 | lr 1.58e-04 | tokens/s 3245 [epoch 0] step 6000 | loss 0.9823 | grad_norm 1.52 | lr 0.42e-04 | tokens/s 3230 ``` ### Checkpoint layout Each checkpoint is saved under `/epoch__step_/`: ``` outputs/eagle3_llama8b/checkpoints/ epoch_0_step_1000/ config.json # draft model config draft_model.pt # draft model weights eagle3_meta.pt # token mapping (selected_token_ids + mask) optimizer.pt # Adam state (for resume) scheduler.pt # LR scheduler state rng/ # distributed RNG state epoch_0_step_2000/ ... LATEST -> epoch_0_step_6250 ``` ### Resume from checkpoint If training is interrupted, resume from the latest checkpoint: ```yaml checkpoint: restore_from: LATEST ``` Or point to a specific checkpoint: ```yaml checkpoint: restore_from: epoch_0_step_3000 ``` *** ## Step 5 — Serve with SGLang The `serve_sglang` helper converts the training checkpoint into an HF/SGLang-compatible format and launches the server in one command. ### Launch the server ```bash 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 ``` On first launch, the helper: 1. Loads `draft_model.pt` and `eagle3_meta.pt` from the checkpoint 2. Rewrites the architecture name for SGLang compatibility (`LlamaEagle3DraftModel` -> `LlamaForCausalLMEagle3`) 3. Exports `model.safetensors` and `speculative_token_map.pt` into a `model/` subdirectory 4. Launches SGLang with the correct speculative decoding flags ### Serving parameters | Flag | Default | Notes | | -------------------- | ------- | ---------------------------------------------------- | | `--algorithm` | — | `EAGLE3` for EAGLE-3 drafters, `EAGLE` for EAGLE-1/2 | | `--num-steps` | 3 | Speculative steps per draft iteration | | `--topk` | 1 | Branching factor for tree search | | `--num-draft-tokens` | 4 | Budget of draft tokens per branch | | `--dtype` | auto | Must match training dtype (e.g., `bfloat16`) | | `--tp-size` | 1 | Tensor parallelism (shards the target model only) | | `--print-only` | — | Inspect the resolved command without launching | Pass extra SGLang flags after `--`: ```bash 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: ```bash curl http://localhost:30000/generate \ -H "Content-Type: application/json" \ -d '{ "text": "Hello, my name is", "sampling_params": {"max_new_tokens": 64} }' ``` Expected output: ```json { "text": "Hello, my name is Sarah and I am a 25-year-old software engineer...", "meta_info": { "prompt_tokens": 6, "completion_tokens": 64, "accept_length_per_step": 3.2 } } ``` 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: ```bash 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: ```json { "choices": [{ "message": { "role": "assistant", "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." } }], "usage": { "prompt_tokens": 14, "completion_tokens": 112 } } ``` *** ## 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 MSE + token cross-entropy loss. ### YAML config Save as `eagle1_llama8b.yaml`: ```yaml recipe: TrainEagle1Recipe dist_env: backend: nccl timeout_minutes: 30 recipe_args: target_model_name_or_path: meta-llama/Llama-3.1-8B-Instruct train_data_path: ./data/perfectblend_renamed val_data_path: null train_split: "train[:200000]" val_split: null output_dir: ./outputs/eagle1_llama8b seq_length: 2048 micro_batch_size: 1 grad_accumulation_steps: 4 num_workers: 4 num_epochs: 1 # EAGLE-1 specific draft_num_hidden_layers: 1 # number of draft transformer layers hidden_loss_weight: 1.0 # MSE loss on hidden states token_loss_weight: 0.1 # cross-entropy loss on tokens freeze_embeddings: true shuffle_seed: 42 log_every_steps: 10 max_grad_norm: 1.0 optimizer: lr: 1.0e-4 betas: [0.9, 0.95] weight_decay: 0.0 checkpoint: enabled: true checkpoint_dir: ./outputs/eagle1_llama8b/checkpoints # Defaults to safetensors + consolidated; can be omitted. model_save_format: safetensors save_consolidated: true ``` ### Launch ```bash torchrun --nproc-per-node=8 \ -m nemo_automodel.recipes.llm.train_eagle1 \ eagle1_llama8b.yaml ``` ### Serve Use `--algorithm EAGLE` (not `EAGLE3`) for EAGLE-1/2 drafters: ```bash 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 | | EAGLE-1 | EAGLE-3 | | ---------------------- | --------------------- | -------------------------- | | **Draft layers** | 1 (configurable) | 2 (with aux fusion) | | **Training objective** | Hidden MSE + token CE | TTT unroll with decay | | **Vocabulary** | Full target vocab | Reduced (e.g., 8k-32k) | | **Training speed** | Faster | Slower (due to TTT unroll) | | **Inference speedup** | Good (2-2.5x) | Better (2.5-3x) | | **Best for** | Quick experiments | Production deployment | *** ## Example Configs Reference | Config | Target | Variant | Notes | | -------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------ | --------- | -------------------------------------------------------- | | [llama\_eagle3\_mvp.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle3/llama_eagle3_mvp.yaml) | Llama 3.2 1B | EAGLE-3 | Quick test, single GPU | | [llama\_eagle3\_mvp\_flash\_attn.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle3/llama_eagle3_mvp_flash_attn.yaml) | Llama 3.2 1B | EAGLE-3 | With FlashAttention-2 | | [llama\_eagle3\_perfectblend.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle3/llama_eagle3_perfectblend.yaml) | Llama 3.1 8B | EAGLE-3 | Production config, 200k samples | | [llama\_eagle3\_1\_perfectblend.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle3_1/llama_eagle3_1_perfectblend.yaml) | Llama 3.1 8B | EAGLE-3.1 | Production config with `fc_norm` + `norm_output` enabled | | [llama\_eagle1\_mvp.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle1/llama_eagle1_mvp.yaml) | Llama 3.2 1B | EAGLE-1 | Quick test, single GPU | | [llama\_eagle2\_mvp.yaml](https://github.com/NVIDIA-NeMo/Automodel/blob/main/examples/speculative/eagle2/llama_eagle2_mvp.yaml) | Llama 3.2 1B | EAGLE-2 | Same as EAGLE-1 (alias) | ## Troubleshooting | Symptom | Fix | | ------------------------------------------ | --------------------------------------------------------------------------------------------------- | | `OutOfMemoryError` during training | Reduce `seq_length` (1024 instead of 2048) or `micro_batch_size` | | Loss stays flat or NaN | Check `max_grad_norm` (default 1.0), reduce `lr` | | SGLang `model not found` error | Ensure `--algorithm` matches the recipe (`EAGLE3` for train\_eagle3, `EAGLE` for train\_eagle1/2) | | `dtype mismatch` at serving | Pass `--dtype bfloat16` to match training precision | | `conversations` vs `messages` column error | Rename the column in your dataset (see Step 2 warning) | | Checkpoint resume fails | Use `restore_from: LATEST` or the exact subdirectory name like `epoch_0_step_1000` | | Low `accept_length_per_step` at serving | Train longer, use more data, or try regenerating answers with the target model (Option B in Step 2) |