Buckets:
Command-line tools
The build puts the tools in out/bin (macOS) or build/bin (Windows, Linux).
Run
| Tool | Use |
|---|---|
nslm-chat |
Run one prompt (K2-Horizon-MoVA or Qwen3.8-Flash-Next). Read Run and serve |
nslm-serve |
OpenAI-compatible HTTP server (K2-Horizon-MoVA or Qwen3.8-Flash-Next) |
Measure
| Tool | Use |
|---|---|
nslm-mova-bench |
Prompt and decode speed, memory and machine state |
nslm-mova-score |
Logits for KLD against a reference |
nslm-mova-gen |
Greedy and sampled outputs for a set of prompts |
nslm-mova-plcheck |
Prompt-lookup decode against plain decode. The two must be identical |
nslm-mova-kbench |
Speed of one kernel at the MoVA shapes. There is one program for each GPU API. The Metal program also selects Metal kernel variants |
nslm-mova-refcheck |
The GPU engine against a C reference forward on the CPU (KL, argmax, router choices) |
nslm-mova-routes |
Router choices for a prompt and a generation, for expert-cache studies |
nslm-qwen4-gen |
Greedy generation of one prompt through the Qwen3.8-Flash-Next CUDA engine, with prompt and decode speed |
nslm-qwen4-kld |
KL divergence, top-1 agreement and perplexity of a Qwen3.8-Flash-Next model against a reference, on one text. --decode N: the last N rows through the decode kernels. --tail-skip R, --cpu-tail N: the model's run with the tail skip or the CPU tail |
nslm-handoff-bench |
CUDA: the time of a GPU-host hand-off inside a CUDA graph, 48 a forward (plan step D4). --work US: the host work of each |
Make models
| Tool | Use |
|---|---|
nslm-moe |
Seed search for experts (--scope gud) and MLA projections (--scope mla). Qwen3.8-Flash-Next experts take --imatrix. --codec tq --tq-k K --tq-l L writes trellis-coded experts (.blkt) on the CPU. Build it with Clang |
nslm-mova-mlacapture |
Activation statistics and X^T X of the MLA projection inputs |
nslm-mova-pack |
Write a model folder |
nslm-qwen4-pack |
Write a Qwen3.8-Flash-Next MLX folder with seed experts, or (--tq) without the routed experts and with trellis banks, from an MLX base folder |
nslm-qwen4-emulate |
Write seed or trellis experts into Unsloth's Q8_0 GGUF shards in place, so that llama.cpp runs them. A map takes each expert from one of several packs. --dense DIR also writes the dense tensors of a model folder |
Read Make a model folder.
Python tools (tools/)
| Tool | Use |
|---|---|
nanoseedlm_k2.py |
The MLX loader for K2-Horizon |
nanoseedlm_qwen4.py |
The MLX loader for Qwen3.8-Flash-Next (mlx-vlm) |
nanoseedlm_seeds.py |
The seed kernels and modules that the MLX loaders share |
qwen4_golden.py |
A tiny random Qwen3.8-Flash-Next model and its mlx-vlm logits, for tests/test_qwen4_ref.c |
qwen4_tiny_attn.py |
The tiny model with strong attention (tests/data/qwen4_tiny_attn), so engine tests see attention bugs |
qwen4_template_golden.py |
Qwen3.8-Flash-Next's chat template rendered by transformers, for tests/test_qwen4_template.c |
qwen4_cache_sim.py |
Replays a routing trace (Q4E_ROUTES) through expert cache policies: LRU, LFU, a hot set and the optimum |
hf_cuda_test.sh |
Build and run CUDA tests on an HF Jobs GPU node |
hf_qwen4_gen.sh |
Run a Qwen3.8-Flash-Next pack through nslm-qwen4-gen on an HF Jobs GPU node |
hf_qwen4_kld.sh |
Score packs with nslm-qwen4-kld against a reference model on wikitext-2, on an HF Jobs GPU node |
hf_llama_kld.sh |
Score Unsloth packs and emulated seed or trellis packs against BF16 in llama.cpp. REF keeps the BF16 reference for later jobs |
gguf_expert_bytes.py |
The routed experts' bytes and bits for each weight of a GGUF pack on the Hub, from its headers only |
page_resident.py |
How much of a pack's files is in the page cache, by kind (mincore). --drop evicts the shards first |
hf_release_pack.sh |
Build the trellis release folders (P3, oQ8e's 8-bit dense modules, the banks) into the bucket, run each twice and score each against BF16, on an HF Jobs GPU node |
qwen4_release_card.py |
Write the release cards, with each pack's KL as shipped from the release job's llama.cpp logs |
hf_release_publish.sh |
Upload the release folders and their cards from the bucket to private Hub repositories (a CPU job with the owner's token) |
hf_qwen4_routes.sh |
Record routing traces of a pack on chat and code text, on an HF Jobs GPU node |
qwen4_tier_map.py |
Write a per-expert tier map from routing traces (most used experts), at random, or by projection |
qwen4_tier_sim.py |
Simulate a high tier that follows the routing: its share of router choices and its loads |
mova_capture.py |
Activation statistics for the expert search, through MLX |
mova_score.py |
Scores in MLX |
mova_export.py |
Inputs for nslm-mova-score, nslm-mova-gen, nslm-mova-bench and route studies |
tokenizer_golden.py, template_golden.py, mova_affine_golden.py |
Golden files for the tests |
serve_bench.py |
Server speed. It waits for an idle GPU and records machine state |
Xet Storage Details
- Size:
- 5.19 kB
- Xet hash:
- 1fdd5c8fcd42d4d4602b0e9c495bb7e5f2b15b2bf47a3fa3434a5b8b706c7176
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.