Skip to content

Tags: kekzl/imp

Tags

v0.19.2

Toggle v0.19.2's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.19.2 (#1050)

Hardening release: KV prefix-cache double-ownership fix (latent silent
corruption under pool pressure), 399 post-launch CUDA error checks, CUDA
graph RAII, cache_control per-breakpoint pins, make asan. Decode neutral
throughout (402.6 tok/s Coder-30B spec-OFF, all A/Bs in PERF_LOG).


Claude-Session: https://claude.ai/code/session_014fTwmujYTCJmRBRr1xoevs

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

v0.19.1

Toggle v0.19.1's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.19.1 (#1043)

Version bump to 0.19.1, CHANGELOG section cut for 2026-07-17, BENCHMARKS
toolchain line updated. check-release.sh: all gates PASS (verify-fast clean
on an idle GPU incl. the perf gates).

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

v0.19.0

Toggle v0.19.0's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.19.0 — bar-1 PPL parity, long-context speculation, FP8-KV…

… auto, 1173 tok/s @16; competitive re-sweep vs llama.cpp b9976 (#985)

- Version 0.19.0; CHANGELOG section cut with highlights.
- Competitive re-sweep (same-day, pp512/tg128, llama.cpp b9976 e3546c794):
  dense GGUF +42-48%, 35B hybrid UD-Q4KM +18%, Gemma-4 MoE +21%; full hero
  matrix appended to docs/scoreboard.tsv. gpt-oss-20b is now a statistical
  tie and non-hero Qwen3-30B GGUF within noise (llama.cpp MXFP4/MoE decode
  caught up) -- documented in BENCHMARKS + tracked in #984.
- README headline numbers refreshed (incl. stale NVFP4 scoreboard values:
  30B MoE now 389-390, 35B hybrid 311, Gemma-4 271).
- GOAL bar-1 trades: gemm.nvfp4_lm_head listed with measured per-model
  numbers (+1.4..+5.0% PPL for +1.9..+6.6% decode, #982); per-model default
  gate stays open in #982.

check-release.sh all gates green locally (incl. verify-fast).


Claude-Session: https://claude.ai/code/session_01VhnpVUzKtwvBcc1GABVdNZ

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

v0.18.1

Toggle v0.18.1's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.18.1 (#953)

Bump VERSION to 0.18.1, move the Unreleased changelog section to 0.18.1
(decode-graph ctx-topology wedge fix #948/#950, fp8_ssm_proj sidecar #949,
audit #6 fixes #941-#945, docs consolidation #946, shared SSE stream driver
+ request builder #951/#952), and update the BENCHMARKS toolchain line.

Also sanitize maintainer home paths from 24 tracked files ($HOME/models /
Path.home() instead of literal /home/<user>/) — the check-release personal-
path gate was red; the leaks accumulated in scripts, skills, and docs since
the gate last ran.


Claude-Session: https://claude.ai/code/session_016P79pmYssX3FQFSNUDK49F

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

v0.18.0

Toggle v0.18.0's commit message
imp v0.18.0

FA2 head_dim=256 default-on (pp4096 +26%) + Qwen3.5-4B-mxfp4 correctness fixes
(VRAM-budgeted MXFP4->FP16 fallback, reasoning-channel reconciliation,
capture-safe narrow-N GEMM) + deterministic-GEMM validation.

v0.17.3

Toggle v0.17.3's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.17.3 — reserve mandatory NVFP4 decode caches before works…

…paces + KV (#928)

Version bump + CHANGELOG (0.17.3 section: #926 VRAM ordering rework,
#927 KV-floor WARN) + BENCHMARKS toolchain line. No code changes.


Claude-Session: https://claude.ai/code/session_016P79pmYssX3FQFSNUDK49F

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

v0.17.2

Toggle v0.17.2's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.17.2 — context-window auto-detection probes (vLLM/llama.c…

…pp/TGI) (#922)

Bump project VERSION 0.17.1 -> 0.17.2, move the #921 context-window probe
feature from [Unreleased] to [0.17.2] in CHANGELOG, and update the
BENCHMARKS current-version line. Server-compatibility release, no perf change.


Claude-Session: https://claude.ai/code/session_01WAXm3AoXmxQ1gkesrp8EsT

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

v0.17.1

Toggle v0.17.1's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.17.1 — adopt C++23 idioms (to_underlying, deducing this, …

…static operator()) (#920)

Patch release capturing the C++23 idiom sweep (#919) and the tool-call test
coverage (#914) that landed after v0.17.0. Behavior-neutral.

- CMakeLists project VERSION 0.17.0 -> 0.17.1 (propagates to IMP_VERSION_STRING).
- CHANGELOG: new [0.17.1] section; fresh empty [Unreleased].
- BENCHMARKS: current toolchain line -> v0.17.1.

v0.17.0

Toggle v0.17.0's commit message
v0.17.0 — C++23 / Ubuntu 26.04 toolchain + accumulated features

v0.16.2

Toggle v0.16.2's commit message

Verified

This commit was created on GitHub.com and signed with GitHub’s verified signature.
release: v0.16.2 — FP4-attention research batch: ThriftAttention prom…

…otion + paged-FP4-QK scaffolds (#846 program closed) (#873)

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>