Commits

Breadcrumbs

History for

onmain

User selector

All users

All time

Commit History

Commits on Mar 15, 2026

• feat(core): add verbose debug logging to longest_token_prefix fast paths

  Show description for 90d46dc
  JamePeng

  committed
  90d46dc

  Copy full SHA for 90d46dc

• Bump version to 0.3.33

  Show description for 6bbc8d2
  JamePeng

  committed
  6bbc8d2

  Copy full SHA for 6bbc8d2

Commits on Mar 14, 2026

• perf(mtmd): optimize media_id masking with bitwise AND

  Show description for bd0b069
  JamePeng

  committed
  bd0b069

  Copy full SHA for bd0b069

• fix(mtmd): prevent multimodal image freeze by injecting deterministic media IDs

  Show description for dd111eb
  JamePeng

  committed
  dd111eb

  Copy full SHA for dd111eb

Commits on Mar 13, 2026

• Update mtmd_cpp API 20260313

  JamePeng

  committed
  afd06ea

  Copy full SHA for afd06ea

• fix(Llama.generate): add explicit fallback context reset and expand `Llama.generate` docstrings

  Show description for 6f81466
  JamePeng

  committed
  6f81466

  Copy full SHA for 6f81466

Commits on Mar 12, 2026

• fix(sampling): prevent memory drift and hallucinations in logits view

  Show description for aa38850
  JamePeng

  committed
  aa38850

  Copy full SHA for aa38850

Commits on Mar 11, 2026

• feat(_ggml): extend ctypes bindings with more ggml constants, enums, and structs

  JamePeng

  committed
  2bb5cd9

  Copy full SHA for 2bb5cd9

• Update llama.cpp API 20260312

  JamePeng

  committed
  c420a2a

  Copy full SHA for c420a2a

• fix(chat_handler): fix tools and function calling in MTMDChatHandler.

  JamePeng

  committed
  8641667

  Copy full SHA for 8641667

Commits on Mar 10, 2026

• perf(cache): optimize LlamaDiskCache I/O and fix LRU behavior

  Show description for a36e8d2
  JamePeng

  committed
  a36e8d2

  Copy full SHA for a36e8d2

• Sync examples : fix empty items in json_schema_to_grammar.py (#19968)

  JamePeng

  committed
  8e8f36f

  Copy full SHA for 8e8f36f

• perf(cache): upgrade `LlamaRAMCache` to O(1) eviction and set `LlamaTrieCache` as default

  Show description for 5c8e056
  JamePeng

  committed
  5c8e056

  Copy full SHA for 5c8e056

Commits on Mar 9, 2026

• fix(core): disable swa_full for non-SWA models (sync llama.cpp upstream #20291)

  Show description for 9acc507
  JamePeng

  committed
  9acc507

  Copy full SHA for 9acc507

• Update llama.cpp API 20260310

  JamePeng

  committed
  955ac33

  Copy full SHA for 955ac33

• fix(chat_format): fix namespace and variable shadowing of llama modules

  Show description for 7efcb2b
  JamePeng

  committed
  7efcb2b

  Copy full SHA for 7efcb2b

• fix(cache): fix namespace shadowing to prevent AttributeError

  Show description for 5e285fe
  JamePeng

  committed
  5e285fe

  Copy full SHA for 5e285fe

Commits on Mar 8, 2026

• feat(MTMDChatHandler): support audio inputs and fix interleaved media ordering

  Show description for 964160d
  JamePeng

  committed
  964160d

  Copy full SHA for 964160d

• Bump version to 0.3.32

  JamePeng

  committed
  e7e1d48

  Copy full SHA for e7e1d48

• fix(sampling): pass seed to sampling context and remove global mutation

  Show description for ad10cfd
  JamePeng

  committed
  ad10cfd

  Copy full SHA for ad10cfd

Commits on Mar 7, 2026

• perf(hybrid): optimize multimodal single-turn and fix KV clear bug

  Show description for fb3072d
  JamePeng

  committed
  fb3072d

  Copy full SHA for fb3072d

• perf(hybrid): prevent expensive array slicing when cache is disabled

  Show description for 850ed2e
  JamePeng

  committed
  850ed2e

  Copy full SHA for 850ed2e

• perf(hybrid): bypass N-1 evaluation split if max_checkpoints is 0

  Show description for 191e334
  JamePeng

  committed
  191e334

  Copy full SHA for 191e334

• perf(hybrid): eliminate PCIe I/O latency for single-turn workflows

  Show description for 7289b54
  JamePeng

  committed
  7289b54

  Copy full SHA for 7289b54

Commits on Mar 6, 2026

• Bump version to 0.3.31

  Show description for 2c191cb
  JamePeng

  committed
  2c191cb

  Copy full SHA for 2c191cb

Commits on Mar 5, 2026

• fix(mtmd): remove OS-level log suppression to expose critical C++ errors

  Show description for 118a1a8
  JamePeng

  committed
  118a1a8

  Copy full SHA for 118a1a8

• fix(hybrid): implement N-1 checkpointing to support 1-token rollbacks

  Show description for c03ce22
  JamePeng

  committed
  c03ce22

  Copy full SHA for c03ce22

Commits on Mar 4, 2026

• refactor(mtmd): introduce omni-modal media pipeline with experimental audio support

  Show description for b342f70
  JamePeng

  committed
  b342f70

  Copy full SHA for b342f70

Commits on Mar 2, 2026

• Bump version to milestone version 0.3.30.

  JamePeng

  committed
  41959f5

  Copy full SHA for 41959f5

Commits on Mar 1, 2026

• update chat handler

  alcoftTAO

  committed
  2258973

  Copy full SHA for 2258973

• Merge branch 'JamePeng:main' into main

  alcoftTAO

  authored
  db6ee16

  Copy full SHA for db6ee16

• refactor(mtmd): redesign multimodal pipeline for concurrent I/O and hybrid state management

  Show description for 5bf6b6a
  JamePeng

  committed
  5bf6b6a

  Copy full SHA for 5bf6b6a

• fix: Correct the mtmd vision check condition bug

  JamePeng

  committed
  22df747

  Copy full SHA for 22df747

• refactor(chat_handler): extract MTMDChatHandler base class and Simplify the complexity of subsequent multimodal adaptation

  Show description for 8b29b88
  JamePeng

  committed
  8b29b88

  Copy full SHA for 8b29b88

• perf(eval): implement adaptive checkpoint intervals for hybrid models

  Show description for 00da436
  JamePeng

  committed
  00da436

  Copy full SHA for 00da436