ReplaySSM: Cache SSM Inputs, Not State (Tri Dao)
Comments on “ReplaySSM: Cache SSM Inputs, Not State (Tri Dao)”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Tri Dao's ReplaySSM changes SSM decoding so that recent inputs are cached and the recurrent state is rebuilt only when the buffer fills, rather than written to HBM every step. The output is mathematically unchanged, and rollback in speculative decoding becomes a buffer pointer move. Reported results in vLLM on B300 with NVFP4 models: up to 1.48x faster standard decoding, and 1.87–1.96x for speculative decoding, where vLLM's current path falls below standard decoding. These figures come from the author's own benchmarks, with no independent source given.
Why it ranked
No. 5 of 93 with a score of 0.69. The panel rated it highest on technical depth (0.96) and new information (0.94), lowest on audience breadth (0.52).
Every answer behind this is in the table under Score.
Score
0.69total under rubric v2.2
- Significance0.63
- Reader interest0.79
- Audience breadth0.52
- Practical relevance0.74
- Substantiated0.87
- Technical depth0.96
- New information0.94
- Counts against: Promotional0.07
- Counts against: Overclaim0.24
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.67 | 0.45 | 0.78 | 0.63 | 0.33 | +0.24 | +0.152 |
| Reader interest | 0.72 | 0.72 | 0.94 | 0.79 | 0.22 | +0.26 | +0.206 |
| Audience breadth | 0.51 | 0.40 | 0.64 | 0.52 | 0.24 | +0.13 | +0.067 |
| Practical relevance | 0.77 | 0.60 | 0.86 | 0.74 | 0.26 | +0.22 | +0.164 |
| Substantiated | 0.93 | 0.80 | 0.88 | 0.87 | 0.13 | +0.05 | +0.044 |
| Technical depth | 0.99 | 0.92 | 0.98 | 0.96 | 0.07 | +0.05 | +0.048 |
| New information | 0.94 | 0.90 | 0.99 | 0.94 | 0.09 | +0.05 | +0.047 |
| Promotional | 0.07 | 0.08 | 0.05 | 0.07 | 0.03 | −0.19 | −0.013 |
| Overclaim | 0.39 | 0.20 | 0.14 | 0.24 | 0.25 | −0.11 | −0.027 |
| Sum of contributions | +0.689 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.689 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-10-25/5/"><img src="https://rubric.hrih.si/badge/2026-10-25/5.svg" alt="Ranked #5 on Rubric, 25 Oct 2026" width="238" height="28"></a>Claim check
Claim check: 8 no source given. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenBenchmark result
ReplaySSM speeds up vLLM's standard decoding by up to 1.48x end-to-end (1.43x on large MoE models).
The article gives no link or citation for it.
The article says “That alone speeds up standard decoding by up to 1.48x (1.43x on large MoE models”
The source says No source to compare with.
No source givenBenchmark result
ReplaySSM speculative decoding delivers 1.87–1.96x speedup over vLLM's standard decoding at the largest batches, while vLLM's existing speculative decoding falls below standard decoding at serving batch sizes.
The article gives no link or citation for it.
The article says “vLLM’s existing implementation falls below standard decoding at serving batch sizes, while ReplaySSM unlocks 1.87–1.96x speedup”
The source says No source to compare with.
No source givenNumber
ReplaySSM supports 3.0–3.3x more concurrent requests than vLLM's speculative path under a fixed memory budget.
The article gives no link or citation for it.
The article says “It also supports 3.0–3.3x more concurrent requests than vLLM’s speculative path under a fixed memory budget”
The source says No source to compare with.
No source givenGeneralization
ReplaySSM is mathematically equivalent to original decoding up to floating-point error and does not change the output.
The article gives no link or citation for it.
The article says “Notably, ReplaySSM is mathematically equivalent to original decoding up to floating-point error. It changes how the state is computed, but not the output”
The source says No source to compare with.
No source givenNumber
The SSM state update kernel remains the primary latency bottleneck in hybrid models up to 100K tokens of context.
The article gives no link or citation for it.
The article says “The SSM state update kernel remains the primary bottleneck up to 100K tokens, which covers a large fraction of real inference workloads”
The source says No source to compare with.
No source givenBenchmark result
At speculative window T=6 on Qwen3.5-122B, vLLM's baseline speculative kernel costs 4.85× the standard decoding kernel, while ReplaySSM costs between 1.27× and 1.72×.
The article gives no link or citation for it.
The article says “At \(T=6\) the baseline kernel costs 4.85× the standard decoding kernel”
The source says No source to compare with.
No source givenNumber
ReplaySSM roughly halves the dominant state memory traffic per head, from 8dn to 4dn.
The article gives no link or citation for it.
The article says “ReplaySSM halves the dominant state traffic from $8dn$ to $4dn$”
The source says No source to compare with.
No source givenGeneralization
Prior SSM speculative decoding methods such as Mamba-in-the-Llama and STree were evaluated only for Mamba-2 and at batch size 1, and still write at least one recurrent state to HBM every step.
The article gives no link or citation for it.
The article says “Both were also evaluated only for Mamba-2 and at batch size 1”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 1 pts, not on front page
Recorded for context only, checked 25 Oct 2026, 05:46 UTC. It played no part in the ranking.