draft-mtp on PQ2_0 fails: Hadamard-latent table 'token_embd.weight' is read without the inverse transform (draft context not covered by PR #169)

未关闭
#190 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
40/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
cpp

调研方向

Start from the draft-context creation path that emits the Hadamard-latent table error and compare it with the main-model path fixed by PR #169; token_embd.weight is the table read without the inverse transform. Read the PQ2_0 dequant/init code and the kvarn2 draft-cache handling, then reproduce with the provided llama-server command. Done means the draft context initializes for the PQ2_0 MTP file and --spec-type draft-mtp runs without that error.

由索引模型根据 Issue 内容生成。

描述

Version

v0.4.8 (ed8f24f17, build 12471)

Problem

--spec-type draft-mtp fails at draft-context init for the PQ2_0 combined MTP file (Ternary-Bonsai-2-27B PQ2_0 + embedded MTP head as blk.64, 7.66 GB):

Hadamard-latent table 'token_embd.weight' is read without the inverse transform → failed to create MTP context

The main model loads and serves the same PQ2_0 file correctly (PR #169 fixed that path — thank you), but the draft context init path reads token_embd.weight without applying the inverse hadamard transform. Same file, same server: plain serving works (48.6 tok/s @32K q4/q4 measured on our 5060 Ti 16 GB), spec arm aborts with the error above.

Additionally, the kvarn2 draft-cache type is not supported for this draft (the other half of why bonsai-mtp-style setups can't run on beellama today).

Config
llama-server -m Ternary-Bonsai-2-27B-PQ2_0-MTP-Q8_0.gguf -ngl 99 -np 1 \
  -c 32768 -ctk q4_0 -ctv q4_0 -fa on -b 2048 -ub 512 \
  --spec-type draft-mtp --spec-draft-n-max 3

(Run on CUDA SM120 / RTX 5060 Ti; the failure is at init, hardware-independent.)

Context

This blocks running MTP speculation on PQ2_0 packs entirely on beellama — our eval (A23, one-binary consolidation test vs the PrismML fork) had to leave the MTP slot on the prism/ProCreations runtime. The fix shape is presumably the same inverse-transform application PR #169 does for the main model, extended to the draft-context read path. Happy to test a fix — the file, hardware, and full gate harness are standing.

主要语言
C++
星标
1.2k
派生
80
平均合并
4 天 20 小时
30 天内合并 PR
2

环境准备

这个项目没有提供开发容器、Dockerfile 或贡献指南,环境需要你自己搭建:先看它的 README,通用步骤见我们的新手贡献指南。

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

Anbeeld/beellama.cpp 的其他 Issue

查看 Anbeeld/beellama.cpp 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。