docs: mistral.rs parity scan — spawn the gaps as phase items - #22
Merged
Conversation
…e knobs automatic Reviewed mistral.rs v0.8.2 against the roadmap and recorded it the way the colibri scan was: one dated section framing the comparison, one tagged item per real gap, and an explicit already-at-parity / non-goals ledger. Parity is read through the North Star — capability parity, delivered as importer/planner decisions, never as flags. Spawned: RoPE scaling + sliding-window (the silent-wrong-answer gap), a Llama-family port, Qwen3-Embedding on the existing qwen3 arch, architecture auto-detection by checkpoint metadata, sampling breadth, batched forward, in-memory prefix KV reuse, and a fleshed-out P10 (LoRA inference, hot-swap, fine-tune loop). P6/P9 spawned nothing — their ISQ/auto-map/FP8-KV/MTP choices map onto items already open, which the scan records as a finding. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
There was a problem hiding this comment.
🟢 Approval recommended
Documentation-only changes with a few minor wording/consistency nits noted for follow-up.
Pull request overview
Adds a new dated “mistral.rs parity scan” to the roadmap, capturing identified capability gaps vs mistral.rs v0.8.2 and recording explicit “already-at-parity” and “non-goals” ledgers in the same style as the prior colibri scan.
Changes:
- Adds new parity-gap items (P2/P3/P5) for RoPE scaling + sliding-window attention, a Llama-family port, Qwen3-Embedding support, architecture auto-detection, sampling breadth, batched forward, and in-memory prefix KV reuse.
- Expands P10 from a single line into a more detailed LoRA/adapters plan (inference, hot-swap, and a fine-tune loop).
- Adds a dedicated “mistral.rs parity scan (2026-08-21)” section and a mistral.rs entry in References.
File summaries
| File | Description |
|---|---|
| ROADMAP.md | Documents the mistral.rs parity scan findings and spawns roadmap items for the identified parity gaps. |
Review details
Suppressed comments (1)
ROADMAP.md:1927
- Terminology is inconsistent with the “North Star” section heading; use the same spelling here for consistency/searchability.
Voxtral — a different product); multi-node distribution (the North-Star machine is one box); and every
- Files reviewed: 1/1 changed files
- Comments generated: 3
- Review effort level: Lite
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
| - [ ] LoRA adapters; an on-device fine-tune loop (Burn supports training) — learn-by-format / | ||
| personalization on-device. | ||
| *(2026-08-21, fleshed out by the mistral.rs scan — was one line.)* | ||
| - [ ] **LoRA adapter inference** *(mistral.rs parity)* — load a HF PEFT adapter |
| dispatches any checkpoint off `config.json`'s `architectures[0]` (HF) or `general.architecture` | ||
| (GGUF) with `--arch` as a rarely-needed override; Mummu *parses* both today but only *dispatches* | ||
| through the curated registry catalog, so a Qwen2/Qwen3/LFM2/OLMoE/MiniLM checkpoint that is not | ||
| one of the 12 catalog entries needs the caller to know its architecture. The North-Star reading |
| live on the dev box: Vulkan advertises SHADER_F16 on the 4070 Ti SUPER, DX12 does not (2026-07-09 probe). | ||
| - colibri — pure-C MoE weight-streaming engine (VRAM/RAM/NVMe hierarchy, MTP speculation, GBNF grammars, | ||
| persistent compressed KV); the 2026-07-30 parity scan's source — https://github.com/JustVugg/colibri | ||
| - mistral.rs v0.8.2 — Candle-based Rust serving stack (~50 archs, ISQ/GGUF/GPTQ/AWQ/FP8, PagedAttention, |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Reviewed mistral.rs v0.8.2 against the roadmap and recorded it the way the colibri scan was: one dated section framing the comparison, one tagged item per real gap, and an explicit already-at-parity / non-goals ledger. Parity is read through the North Star — capability parity, delivered as importer/planner decisions, never as flags.
Spawned items: RoPE scaling + sliding-window attention (the silent-wrong-answer gap), a Llama-family port, Qwen3-Embedding on the existing qwen3 arch, architecture auto-detection by checkpoint metadata (P3), sampling breadth / batched forward / in-memory prefix KV reuse (P5), and a fleshed-out P10 (LoRA inference, hot-swap, fine-tune loop). P6/P9 spawned nothing — their ISQ / auto-map / FP8-KV / MTP choices map onto items already open, which the scan records as a finding.
🤖 Generated with Claude Code