gguf: parser for percentage-mixed GGUF filenames

[mishig25]

Adds a parser for percentage-mixed GGUF filenames — files that ship a single artifact with tensors quantized to multiple ggml types and encode the per-type byte share in the name as <pct><quant> tokens.

Example: DeepSeek-V4-Flash-55IQ2_XXS-34Q2_K-07Q8_0-03F16.gguf (antirez/deepseek-v4-gguf).

API

parseGGUFQuantMix("DeepSeek-V4-Flash-55IQ2_XXS-34Q2_K-07Q8_0-03F16.gguf")
// {
//   components: [{pct:55,quant:'IQ2_XXS'}, {pct:34,quant:'Q2_K'}, {pct:7,quant:'Q8_0'}, {pct:3,quant:'F16'}],
//   dominant:   {pct:55, quant:'IQ2_XXS'},
// }

Also exports GGUF_QUANT_MIX_COMPONENT_RE and GGUFQuantMix / GGUFQuantMixComponent types.

parseGGUFQuantLabel delegates to it

When the filename looks like a mix, parseGGUFQuantLabel now returns the dominant component instead of the file-order last match (which today surfaces the smallest tail — F16 in the example above). Plain single-quant filenames are unchanged.

parseGGUFQuantLabel("DeepSeek-V4-Flash-55IQ2_XXS-34Q2_K-07Q8_0-03F16.gguf") // → "IQ2_XXS"
parseGGUFQuantLabel("Meta-Llama-3.1-8B-Instruct-Q4_K_M.gguf")               // → "Q4_K_M"  (unchanged)
parseGGUFQuantLabel("Qwen3-4B-UD-Q2_K_XL.gguf")                             // → "UD-Q2_K_XL"  (unchanged)

Notes

• Returns undefined for plain single-quant filenames so it composes cleanly with the existing fallback.
• Delimited lookbehind / lookahead avoids false positives on size labels like 7B.
• Quant alternation is length-sorted so IQ2_XXS wins over the prefix IQ2_XS.

---

Note

Low Risk
Low risk: adds a new filename parser and adjusts parseGGUFQuantLabel output only for percentage-mixed names, with extensive unit test coverage; no GGUF binary parsing or data handling logic is changed.

Overview
Adds support for percentage-mixed GGUF filenames (e.g. Model-55IQ2_XXS-34Q2_K-07Q8_0-03F16.gguf) by introducing parseGGUFQuantMix, GGUF_QUANT_MIX_COMPONENT_RE, and the GGUFQuantMix* types in @huggingface/tasks, and re-exporting them from packages/gguf.

Updates parseGGUFQuantLabel to detect these mixed names and return the dominant (largest %) quant instead of the last match, and adds comprehensive tests covering suffixes like -imatrix/-MTP, path prefixes, size-label false positives, and non-quant storage types.

^{Reviewed by Cursor Bugbot for commit f4a6f75. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit a6e16c8. Configure here.}

[pcuenca]

Looks good based on the tests