aswinkumar99/qwen3.8-flash-next-expert-traces

Qwen3.8-Flash-Next expert routing traces

Token-level routing traces of a deployed MoE model: for every token and every one of the 48 MoE layers, which experts the router chose, the top-32 router logits behind that choice, and the exact hidden state the router read — plus, in v3, the state at many layers per token, the post-final-norm state the LM head consumes, and the LM head's top-8 next-token candidates.

The corpus exists to answer one question: how well can the next tokens' expert routing be predicted in advance, so an engine that cannot hold all experts in VRAM can prefetch them instead of stalling. It is equally usable for studying MoE routing itself.

Model

checkpoint	RadixArk/Qwen3.8-Flash-Next-NVFP4
architecture	qwen4_exp, 48 layers, 512 experts/layer, top-10, hidden 2560, MoE intermediate 640
extras	4 hyper-connection streams (hc_count=4, hc_lowrank=320), PLE n-gram table (16 heads), untied embeddings, vocab 248,320
quantisation	NVFP4 (modelopt 0.46.0), group size 16, routed experts only
routers are NOT quantised	*.mlp.gate* is in the quant ignore list, so every gate stays bf16 — and carries no bias. The router is exactly W_l @ x.
engine	FreeToken 0.1.2, branch expert-trace-study

Sampling for every run unless a run.json says otherwise: temperature 0.2, top_p 1.0, top_k disabled. The model produced its own continuations, so the traces are on-distribution.

Layout

traces/<run_id>/trace.part-*.bin      # the binary trace (format below)
traces/<run_id>/prompts.jsonl         # prompts, subject to the policy below
traces/<run_id>/run.json              # task id, category, seed, sampling params, sizes, engine state
traces/<run_id>/server.log
diagnostic/<task>/                    # the richer all-48-tap subset
diagnostic/router_weights.npz         # every layer's gate matrix, fp32
manifest.jsonl                        # one line per run
docs/trace_format.md                  # the authoritative format description
RECORDER_NOTES.md                     # what changed vs v2, how to read v3, known caveats

Format

docs/trace_format.md is authoritative and is copied verbatim from the engine repo. In brief:

• Little-endian, append-only: 32-byte file header, JSON run header, then fixed-size token blocks.
• v2 = the original five state probes {0,12,24,36,47} in float16.
• v3 = a configurable probe list (up to all 48), an optional post-final-norm probe (layer id 255), bfloat16 storage, LM-head candidates, and per-layer device timestamps. A file is only v3 if it uses one of those; a v2 reader refuses v3 rather than misreading it.

Read a trace with the engine repo's tools/expert_trace/read.py:

from read import open_trace
tr = open_trace("trace.part-000.bin", load_states=True)
tr.steps                       # one record per token
tr.records                     # [tokens, layers] routing + top-32 logits
tr.state_float("block_input")  # [n, 2560] router inputs, float32 (widens bf16)
tr.state_probe_layers          # probe order within a token; 255 = final-norm state
tr.candidates                  # top-8 next-token ids + probabilities, per decode row
tr.layer_times                 # device-clock ns per (token, layer, {attn, router, moe})

Prompt policy

Trace token ids are always present, so teacher forcing works for every run regardless. Prompt text follows the source's licence:

source	prompt text	why
the study's own agentic tasks, MATH, GSM8K, OSS-Instruct/Magicoder	full text	permissive
LMSYS-Chat-1M, WildChat	hash + dataset id + row index only	gated / redistribution-restricted
IETF RFCs, Project Gutenberg texts (long-corpus real-document tasks)	full text, incl. the Project Gutenberg header/footer that each file keeps (the letter of its licence)	RFC Editor terms allow verbatim reproduction; Gutenberg texts are public domain in the USA; sources, SHA-256 and token counts in workloads/docs/*/MANIFEST.md

Each manifest line carries prompt_text_included: true|false.

Two things to read before using the cache fields

1. hit_mask and slots_in_use here are not deployment cache behaviour. These runs keep all 24,576 experts resident, so the LRU never evicts. Every miss is a cold-start miss — the first touch of an expert — and the rate falls to zero as coverage grows. Cache and timing data for a real offload deployment must come from a run that actually offloads.
2. globaltimer is not step_wall_ns. Per-layer stamps are CUDA device nanoseconds, monotonic within a run and not comparable across GPUs. The run header carries one {globaltimer_ns, clock_realtime_ns} pair to align the two axes.

Licence and attribution

Traces and manifests: CC-BY-4.0.

• Model: Qwen3.8-Flash-Next (Apache-2.0); NVFP4 quantisation by RadixArk.
• Engine: FreeToken (Apache-2.0).
• Prompt sources are credited per run in manifest.jsonl, each under its own licence.

Generated text carries the usual caveat: it is model output at temperature 0.2 and was not reviewed for accuracy.