MakazhanAlpamys/Soup v0.71.23 on GitHub

What's New

Native Spectrum targeted training (closes #266). Fine-tune only the layers that matter — and find them on a CPU box.

soup spectrum scan --model <id|path> --top-percent 50 [--modules mlp,attn|all] [-o patch.yaml] streams a model's .safetensors shards one tensor at a time (no model load — peak RAM is the largest single weight matrix) and computes a singular-value SNR per weight matrix with a Marchenko-Pastur noise threshold (arXiv:2406.06623). It prints a per-group SNR table and a ready-to-paste training.unfrozen_parameters YAML block; results cache under ~/.soup/spectrum/.
Targeted full fine-tuning — with training.unfrozen_parameters set, soup train freezes every parameter and unfreezes only the matched high-SNR layers (full FT, LoRA off).
The SNR kernel is pure-numpy and transpose-invariant; GPT-2 Conv1D naming (c_attn/c_fc/c_proj) is recognised alongside Llama-style names. The existing spectrum trainer-plugin wrapper is untouched. LISA (per-step layer sampling) is the follow-up in #267.

Live-validated on Windows + RTX 3050: CPU scan of SmolLM2-135M + a real top-25% unfrozen full-FT train (loss 3.455 → 0.719).

pip install -U soup-cli
# scanning a Hub model id needs torch for bf16 weights:
pip install -U 'soup-cli[train]'

unfrozen_parameters patterns are validated at parse time: nested-unbounded-quantifier regexes (ReDoS) are rejected, along with null bytes and empties; count (50k) and length (512) are capped.
Hub downloads route through the SSRF-hardened, namespace-pinned loader; symlinked shards and matrices above a 2^31-element SVD cap are skipped; --output stays under cwd.
A real grad-checkpointing backward-pass bug (frozen embeddings) and the ReDoS hole were both caught in review and fixed before release.

Very large models need RAM ≈ the largest scanned weight matrix. The per-tensor stream keeps peak RSS to one matrix, but a 70B's biggest mlp/attn weight is still multi-GB; use --modules mlp,attn to skip the giant embedding/lm_head matrices (which also exceed the 2^31-element SVD cap and are skipped under all).
bf16 weights require torch — the pure-numpy framework cannot read bf16; without soup-cli[train] the scan raises a friendly error naming the install.
Spectrum is full fine-tuning — unfrozen_parameters requires quantization: none and is text / transformers / task=sft only; QLoRA users keep LoRA and drop unfrozen_parameters.
A true 70B CPU scan is validated on N=0 (the streaming/cap/cache logic is unit-tested; the live scan was validated on SmolLM2-135M) — same infra class as the multi-GPU items.

Tests: 14184 → 14278 (+94).