weifanjiang/spec-decode-profilings

Speculative decoding profiles

Per-step timing and acceptance of speculative decoding in vLLM under load, recorded with sd-profiler. Each run's model, drafter checkpoint, GPU and software versions are in the runs table.

How it was measured

• Engine: vLLM, greedy (temperature 0), one GPU per run.
• Load: specdec_bench keeps C requests in flight on a SPEED-Bench split (first turns, up to 512 new tokens), with EAGLE-4's workload sizes: 22, 88, 128, 220, 440 and 880 prompts at C = 1, 8, 32, 64, 128 and 256, and four prompts per request in flight (at least 22) at other C. Each timed pass follows a matching warm pass on the same engine; only timed passes are here.
• Splits: qualitative (880 prompts in 11 categories) and the throughput splits throughput_1k ... throughput_32k (1,536 prompts each, all of one length, counted with OpenAI's o200k tokenizer; three categories: low, mixed and high entropy).
• Accuracy runs: one pass with the token log on (prompts in runs gives the count). Drafters keep their draft logits (vLLM's "probabilistic" drafting, which at temperature 0 still picks the top token). These runs are slower and are not meant for timing.
• Step log: a vLLM plugin records CUDA events around every engine step.

Tables

Paths are TABLE/GPU/TARGET/SPLIT/DRAFTER-kK/.... Every file carries gpu, target, split, drafter, k, c and start (repeat number; 0 for accuracy runs), so files stack.

runs: one row per run, mode = load or accuracy: throughput (throughput_tps, generated tokens ÷ elapsed workload time), output tokens, step counts, guessed and committed tokens, versions and model revisions.

steps: one row per engine step of the timed pass.

Column	Meaning
wall	Unix time when the step entered the model runner
num_reqs, num_tokens	Requests and tokens in the step
decode_reqs, spec_tokens	Requests decoding, and guesses they carry
prefill_tokens	Tokens of requests still in prefill
ctx_tokens	Tokens already in the KV cache, summed over requests
full_graph, padded_tokens	Whether the target ran as one CUDA graph; tokens after padding
hist	Requests by tokens committed this step (index 0–16: accepted guesses plus one)
t_fwd, t_fwd_end	GPU ms from the step's start to the target's forward start and end
t_lg0, t_lg1	Same, for the target's LM head
t_samp, t_end	Same, for the start of sampling/verification and the end of the step
t_d0, t_d1	Same, for the drafter's propose call
since_prev_ms	GPU ms from the previous step's start

requests (load runs): one row per request: question_id, category, prompt_tokens, streamed chunk sizes and times (chunks, times), and the generated token ids (output_ids).

tokens (accuracy runs): one row per guessed token.

Column	Meaning
step, wall	Verification step counter and Unix time
req	Request index within the run (matches req in outputs)
batch	Requests in the step
draft_pos	Position of the guess in its draft, 1..k
ctx, out_pos	Tokens before the guess; its position in the output (0-based)
token, target_top	The guess and the target's top token there
accepted	Whether the guess was kept
t_prob, t_entropy, t_margin	Target's probability of the guess, entropy (nats), top-1 minus top-2 probability
d_prob, d_entropy, d_margin	Same for the drafter (NaN for n-gram, which has no probabilities)

outputs (accuracy runs): each answer's output_ids, with req, question_id, category and prompt_tokens, for comparing outputs against plain decoding.

Code

Recorded with sd-profiler: scripts/run_load.py (driver), sd_profiler/ (the vLLM plugin with the step and token logs), scripts/to_hf.py (this layout).