Feature/spec decode draft model

[tomasruizt]

Purpose

Enabling draft models for speculative decoding (SD).
E.g. Qwen3-1.7B as draft model and Qwen3-32B as target model.
This type of SD requires no special trained heads (like EAGLE, or Medusa).

Example usage:

vllm serve \
    --model=Qwen/Qwen3-4B \
    --speculative-config '{"model": "Qwen/Qwen3-0.6B", "method": "draft_model", "num_speculative_tokens": 3, "max-model-len": 2000}' \
    --max-model-len 2000

Get a generation:

curl -X POST http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{"prompt": "Capital of France", "max_tokens": 16}'

Status

• The acceptance rates in greedy decoding are great for Qwen3 models (see corresponding section).
• Using SD with Qwen3 has higher throughput (TPOT) than not using SD.

Acceptance Length

As suggested by @ekagra-ranjan, I benchmarked acceptance length (AL) with the command below:

VLLM_USE_V1=1 python3 examples/offline_inference/spec_decode.py \
    --model-dir Qwen/Qwen3-32B \
    --draft-model Qwen/Qwen3-1.7B \
    --method draft_model \
    --num_spec_tokens 3 \
    --dataset-name hf \
    --dataset-path philschmid/mt-bench \
    --num_prompts 100 \
    --temp 1.0 \
    --gpu-memory-utilization 0.9

The AL values within the Qwen3 family seem good, both with temperatures of 0.0 (greedy) and 1.0.
As a sanity check, I benchmarked LLama-3.2-1B as both target and draft, which had almost perfect AL (3.97/4), suggesting its working as intended.
I have not run the default model meta-llama/Llama-3.1-8B-Instruct, because I didn't find a good draft model for it, but feel free to suggest one and I can run the benchmarks.

Temperature t=0:

Target	Draft	K	Temperature	AL
Qwen3-32B	Qwen3-1.7B	3	0.0	2.79
Qwen3-32B	Qwen3-1.7B	4	0.0	3.12
Llama-3.2-1B	Llama-3.2-1B	3	0.0	3.97

Temperature t=1.0:

Target	Draft	K	Temperature	AL
Qwen3-32B	Qwen3-1.7B	3	1.0	2.61
Qwen3-32B	Qwen3-1.7B	4	1.0	2.85
Llama-3.2-1B	Llama-3.2-1B	3	1.0	2.82

Using t=1.0, the AL metric degrades. However, spec-decode with probabilities is not yet implemented, needed for lossless rejection sampling. This is being addressed atm: #20459. After that PR is merged, the AL for non-greedy spec-decode should improve.

All scripts and logs used for the benchmarks can be found in this Google Drive.

Online Throughput Metrics

I measured online throughput metrics using the commands below. After making sure the draft model also uses CUDA graph, SD has higher throughput than not using SD. See tables below.

VLLM_USE_V1=1 vllm serve Qwen/Qwen3-32B \
  --max-model-len 20000 \
  --disable-uvicorn-access-log

# or 

VLLM_USE_V1=1 vllm serve Qwen/Qwen3-32B \
  --speculative_config '{"method": "draft_model", "model": "Qwen/Qwen3-1.7B", "num_speculative_tokens": 3, "max_model_len": 20000}' \
  --max-model-len 20000 \
  --disable-uvicorn-access-log

nohup vllm bench serve \
  --model Qwen/Qwen3-32B \
  --endpoint-type openai-chat \
  --endpoint /v1/chat/completions \
  --dataset-name hf \
  --dataset-path philschmid/mt-bench \
  --num-prompts 80 \
  --max-concurrency (100|1) \
  --temperature 0.0 \
  --top-p 1.0 2>&1 > results/qwen3-32b-t0-run1.out &

• The tables show shorter runtime and higher throughput for SD (both in batch size 1 and 100).
• Using SD the TPOT is 50% shorter (better) in batch size 1, and 26% to 33% shorter in batch size 100. The reason is the higher throughput of the draft model.
• Using SD the TTFT and ITL are higher (worse), because tokens are produced in batches by the spec-decoding. Nevertheless, total runtimes are shorter overall when using SD.

The metrics (lower is better) are:

• TTFT: Time-to-first-token
• TPOT: Time-per-output-token
• ITL: Inter-token-latency

Batch Size = 1

For Temperature = 0.0:

Target	Draft	Runtime	TTFT	TPOT	ITL
Qwen3-32B	-	943s	69.14ms	46.06ms	45.88ms
Qwen3-32B	Qwen3-1.7B	466s	76.77ms	22.62ms	62.44ms

Using SD runtimes and TPOT are shorter by ~50%.

Batch Size = 100

For Temperature = 0.0:

Target	Draft	Runtime	TTFT	TPOT	ITL
Qwen3-32B	-	16.88s	262.59ms	65.09ms	64.84ms
Qwen3-32B	Qwen3-1.7B	13.04s	284.45ms	43.70ms	121.48ms

For Temperature = 1.0:

Target	Draft	Runtime	TTFT	TPOT	ITL
Qwen3-32B	-	16.83s	230.04ms	64.95ms	64.70ms
Qwen3-32B	Qwen3-1.7B	14.84s	272.51ms	48.00ms	122.27ms

This scenario with batch size 100 is a more realistic inference case.
Using SD runtimes and TPOT are shorter.

Profiling

This section was removed, since using CUDA graphs on the draft model significantly improved its speed.

Profiling script

I used the command below to profile the generation process and identify that the draft model was running too slow before.

export VLLM_USE_V1=1
export VLLM_TORCH_PROFILER_DIR=./profiles/
export CUDA_LAUNCH_BLOCKING=1

vllm bench throughput \
    --model=Qwen/Qwen3-32B \
    --speculative-config '{"model": "Qwen/Qwen3-1.7B", "method": "draft_model", "num_speculative_tokens": 3, "max_model_len": 2048}' \
    --dataset-name=hf \
    --dataset-path=likaixin/InstructCoder \
    --max-num-seqs=100 \
    --num-prompts=10 \
    --input-len=1000 \
    --output-len=10 \
    --max-model-len=2048 \
    --gpu-memory-utilization=0.95 \
    --profile

Note: The command uses the --profile flag, which I introduce in this PR: #24575

Test Plan

The added unit test check the correctness metrics. To run it:

cd tests/v1/e2e/
pytest test_spec_decode.py -k test_draft_model_correctness

EAGLE testing

I tested that the EAGLE implementation stays unaffected the command below

VLLM_USE_V1=1 python3 examples/offline_inference/spec_decode.py \
    --model-dir meta-llama/Llama-3.1-8B-Instruct \
    --eagle-dir yuhuili/EAGLE3-LLaMA3.1-Instruct-8B \
    --method eagle3 \
    --num_spec_tokens 7 \
    --dataset-name hf \
    --dataset-path philschmid/mt-bench \
    --num_prompts 80 \
    --temp 0.0 \
    --gpu-memory-utilization 0.9

The results are in line with previous measurements like #17504 (comment)

total_num_output_tokens: 16990
num_drafts: 4816
num_draft_tokens: 33712
num_accepted_tokens: 12208
mean acceptance length: 3.53
--------------------------------------------------
acceptance at token 0: 0.74
acceptance at token 1: 0.54
acceptance at token 2: 0.41
acceptance at token 3: 0.31
acceptance at token 4: 0.23
acceptance at token 5: 0.17
acceptance at token 6: 0.13

Follow-up Optimizations

• Include the tokens in next_token_ids together with target_token_ids in the first forward pass of the draft model. This reduces the number of forward passes needed in each drafting phase by one, speeding up drafting.

(Optional) Documentation Update

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors.

You ask your reviewers to trigger select CI tests on top of fastcheck CI.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

🚀

[gemini-code-assist]

Code Review

This pull request introduces support for speculative decoding using a draft model. The changes are comprehensive, touching configuration, model loading, scheduling, and the core speculative decoding logic. New tests and benchmark modifications are also included to validate and measure the new feature. The overall implementation appears solid. However, I've identified a critical issue in a refactoring of the bind_kv_cache utility function, which removes an important safety check and could lead to incorrect behavior for certain model architectures.

[ekagra-ranjan]

@tomasruizt - Thank you for the PR!

1. Can you also report Acceptance Length (AL) and K used? It is more informative than AR. For e.g., with Eagle with K=3 we get AL ~2.29 on MTBench so we can expect 2.29x speedup assuming 0 draft overhead. AL gives a more holistic overview of the speedup for a given K than AR.
2. Can you run the metrics on MTBench instead of one off sample for less noisy metric? This would also allow us to compare the AL and TPOT improvement from other SD methods: [Benchmark][V1][Spec Decode][EAGLE] Tracking benchmark for V1 EAGLE #17812. You can find the cmd to run the offline inference to find AL and online inference to find TPOT using MTBench in vLLM here: [V1][Spec Decode][Feature] Spec decode with probs #20459. You will need to update offline inference script to use separate draft model.

[ekagra-ranjan]

As a main model Qwen3-1.7B runs a decoding forward pass in 9-10ms, while when its used as a draft model, each forward pass takes 19-22ms.

What is the TP you are using for Qwen3-32B? By default, draft model TP is equal to target model TP. Since Qwen3-1.7B is a small model, running it on high TP might be incurring nccl communication cost. Try setting draft TP to 1.

[tomasruizt]

What is the TP you are using for Qwen3-32B? By default, draft model TP is equal to target model TP. Since Qwen3-1.7B is a small model, running it on high TP might be incurring nccl communication cost. Try setting draft TP to 1.

I ran the benchmarks with TP=1 and num_draft_tokens=3. So we can rule out TP communication issues.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @tomasruizt.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[tomasruizt]

@ekagra-ranjan I updated the metrics with your suggestions. Acceptance lengths look good IMO. After applying CUDA graph on the draft model its speed improved dramatically. Now the throughput of using SD is higher than not using it (in particular TPOT).

[llsj14]

So maybe this file name should be changed.

[tomasruizt]

Maybe. I feel like a refactoring is needed to cleanly separate EAGLE from "draft_model", to simplify the if-else statements and reduce the mental overhead when reading the code.

[tomasruizt]

But perhaps that should be a follow-up PR to avoid changing too much code in a single PR.

[ekagra-ranjan]

@tomasruizt - The online inference cmd shared in PR has this flag --max-concurrency 100. This means that you are testing on Batch size of 100. Please use 1 instead of 100 to iterate as of now. When things are finalized then you can provide numbers for BS=1/4/16.

Also, please provide the cmd you used to profile the model for reproducibility.

[tomasruizt]

@ekagra-ranjan

• I've added throughput metrics with batch size = 1.
  • With batch_size=1, using SD reduces runtime and TPOT by 50%. I.e. inference is 2x as fast 🚀
  • With batch_size=100, using SD also reduces runtime and TPOT, by 26% to 33%. This batch size gives a better impression of SD performance under a more realistic and challenging workload.
• I've added the profiling command in the Profiling section

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @tomasruizt.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork