---
license: apache-2.0
task_categories:
- other
tags:
- mixture-of-experts
- expert-routing
- qwen3
- inference-optimization
size_categories:
- 100M<n<1B
configs:
- config_name: default
  data_files: layers_v1/*/labels/*.safetensors
---

# CommitMoE — Qwen3.5-35B-A3B-FP8 expert-routing traces, columnar by layer

Per-token MoE routing decisions from `Qwen/Qwen3.5-35B-A3B-FP8`, laid out
**one directory per layer** so a predictor for a single layer reads only what it
needs instead of scanning interleaved shards.

The model has **40 MoE layers, 256 experts each, top-8 routing**, hidden size
2048. Every row is one `(prompt, decode token, layer)` triple.

## What this is for

Predicting *which experts a layer will route to* a couple of layers ahead, so
their weights can be prefetched over PCIe into a small GPU-resident cache
before the layer runs. A prediction that misses costs a stall; the metric that
matters is recall at the cache size `C`, and under a `W = C` admission policy
the stall count per token per layer is exactly `8 · (1 − recall@C)`.

## Layout

```
layers_v1/<chunk>/
  prompts.jsonl                       # id + prompt_idx, one row per prompt
  labels/part_*.safetensors           # top_indices[8], prompt_idx, tok, layer, gen_token_id
  hidden/layer=NN/part_*.safetensors  # h[2048] (pre-MoE hidden), prompt_idx, tok
  logits/layer=NN/part_*.safetensors  # logits[256] (router), prompt_idx, tok
```

`labels` covers all 40 layers in one table — it is small (~2.6 GB per 10k
prompts) and window features read every layer of it. `hidden` is the expensive
part at about **10.0 GB per layer per 10k prompts**; `logits` is **1.27 GB**.

## Chunks and prompt ids

`prompt_idx` is **globally unique across chunks**, so they can be pooled
without collision:

| chunk | prompts | `prompt_idx` range |
|---|---|---|
| `chunk_10k` | 10,000 | 0 – 9,999 |
| `chunk00` | 4,000 | 10,000 – 13,999 |
| `chunk01` | 6,000 | 14,000 – 19,999 |

`chunk00` lives in the companion repo
[`commitmoe-qwen35-fp8-layers`](https://huggingface.co/datasets/RASMUS/commitmoe-qwen35-fp8-layers).
Each chunk's `prompts.jsonl` starts at its own base, and readers derive that
base from the file's first row rather than assuming zero.

**A note on the ids, because it is easy to get wrong.** The raw shards store
`prompt_idx` as a *shard-local slot* (0–31, an index into that shard's own
32-entry prompt list), not a global id. Writing it through unchanged collapses
every shard's prompts onto the same 32 ids — silently, since the output still
looks well-formed. The files here are built by resolving each shard's prompt-id
*strings* through the chunk's index, and every upload is checked by comparing
the written `prompt_idx` set against the set resolved from the shards.

## How the shards were laid out upstream

Each raw shard is a **decode-step slice** of a 32-prompt batch — 32 prompts ×
~26 decode steps × 40 layers ≈ 32,768 rows — so roughly **ten consecutive raw
shards share the same 32 prompts**. That is invisible in this columnar form but
matters if you rebuild from the raw traces.

## Splits

Not stored. The split is derived from the prompt-id string, so it stays
consistent across chunks and rebuilds.

## Related

- Raw interleaved traces: [`commitmoe-qwen35-fp8-expert-routing-traces`](https://huggingface.co/datasets/RASMUS/commitmoe-qwen35-fp8-expert-routing-traces)
- `chunk00` + raw for `chunk00`/`chunk01`: [`commitmoe-qwen35-fp8-layers`](https://huggingface.co/datasets/RASMUS/commitmoe-qwen35-fp8-layers)