repowiserepowise
Sign in
rakaarwaky/hipfire
OverviewDocsArchitectureKnowledge GraphFilesCode HealthRefactoring

People & History

CommitsContributorsDecisions
ChatPro
Stats
repowiserepowise
ExplorePricingDocs
Sign inIndex repoIndex your repo free
repowiserakaarwaky/hipfire

Commits

Every commit scored for change-risk against this repo's own history, so 'elevated' means elevated here rather than on some global curve.

Needs review

1,026of 3,072 scored

1,026 commits sit in this repo's top risk tercile, which is 33% of the 3,072scored. The cut is drawn against this codebase's own history rather than a global curve, so a quiet repo still fills its top band, and here it starts at 8.4 out of 10. What pushes a commit up is size and spread together: a large change confined to one area scores below a smaller one scattered across a dozen files.

Fix commits65221%Commits whose subject reads as a bug fix rather than new work.Written by an agent2,13670%Read from commit trailers, so it counts what was declared.Change diffusion0.82bitsShannon entropy of a commit's churn across its files. Zero is a single file, and every extra bit is a doubling of how widely the change spread.Review threshold8.4out of 10Score a commit has to clear to land in this repo's top tercile.

How the work changed shape

Commit categories over time, read off the subject line. Fixes carry the accent because that is the series this chart exists to show.

Consistently other-driven across its history.

Feature26%
Fix22%
Refactor15%
Docs12%
Test3%
Deps0%
Chore4%
Other17%
70%of indexed commits are agent-attributed(2136 of 3072)
claude (2082) · opencode (51) · codex (3)

Review-priority queue

Ranked by change-risk, highest first. Priority is a tercile of this repo's own distribution, so a quiet repo still fills its top band.

Commit review-priority queue
#CommitAuthorWhenLinesRiskTop driver
1
37b36c8brefactor(loading): collapse qwen35 to one public load_weights entry
Björn Bösel
1mo ago+17.8K -5.8K
100%Elevated
more lines added than baseline
2
09800fbfrefactor(rdna-compute): decompose GPU dispatch object (#342)
Björn Bösel
2mo ago+35.2K -34.2K
100%Elevated
more lines added than baseline
3
a18d694achore(quality-eval): generate eval slice + populate canary fixture
Kevin Readclaudeassisted
3mo ago+45.8K -31
100%Elevated
more lines added than baseline
4
a020c8e1feat(cdna): MI300X calibration optimization (Phase 2-5) + gfx942 baseline + wave64 residual gemv (#90)
Kaden Schuttclaudeassisted
3mo ago+65.4K -48
100%Elevated
more lines added than baseline
5
bccc9a17wmma correctness fix + MQ6 family + cross-arch prefill + gate framework
Kaden Schuttclaudeassisted
4mo ago+35.4K -11.8K
100%Elevated
more lines added than baseline
6
72620fc9Track C: HFQ4 KV cache — 5 experiments completed
Kaden Schuttclaudeassisted
4mo ago+72.3K -56
100%Elevated
more lines added than baseline
7
a9dc252fPer-layer profiling infrastructure + full benchmark report
Kaden Schuttclaudeassisted
4mo ago+482.6K -0
100%Elevated
more lines added than baseline
8
94d0bf2bfeat(runtime): FireMap - grow Qwen3.5 KV on demand with HIP VMM (#549)
黃思銘-HKUST
1w ago+8.1K -746
99%Elevated
more lines added than baseline
9
a064932cfeat(kernels): lower radiowave into beta and compile kernels through it
Kaden Schutt
2w ago+9.0K -20
99%Elevated
more lines added than baseline
10
64f16d22docs(skills): promote canonical workflow overhaul
Kaden Schutt
3w ago+4.0K -3.2K
99%Elevated
more lines added than baseline
11
1a0acd11docs: promote lfm-redline documentation foundations
Kaden Schutt
3w ago+8.5K -4.9K
99%Elevated
more lines added than baseline
12
3240ad0edispatch: graft redline replay behind fail-closed gate
Kaden Schutt
4w ago+12.7K -0
99%Elevated
more lines added than baseline
13
53bdf083feat: agentic PR merge-gate (Tier-3 GPU gate + autoresearch/ar package)
Kaden Schuttclaudeassisted
4w ago+11.3K -0
99%Elevated
more lines added than baseline
14
1876bd8efix(gate-up): make nosync variants buildable and opt-in
Kaden Schutt
2mo ago+5.8K -2.3K
99%Elevated
more lines added than baseline
15
0be4ad16refactor(paro): remove stale env-gated experiments
Björn Bösel
2mo ago+15.0K -7.3K
99%Elevated
more lines added than baseline
16
0884535cQwen3.5/3.6 grammar-guided tool calling and prompt caching
Nick Woolmer
2mo ago+6.5K -234
99%Elevated
more lines added than baseline
17
1cafa192[codex] Land A3B MTP and DFlash perfmaxx work (#352)
Kaden Schutt
2mo ago+5.6K -1.5K
99%Elevated
more lines added than baseline
18
15a96b98research(paroquant): add G256 milestone probes
Kaden Schutt
2mo ago+6.5K -676
99%Elevated
more lines added than baseline
19
b263fb3cfeat(arch-deepseek4): add DeepSeek V4 Flash support (arch_id=9)
Nick Woolmerclaudeassisted
2mo ago+25.3K -3.1K
99%Elevated
more lines added than baseline
20
6c014fc6dots-ocr: phase 0 item 5 — HF + vLLM reference captures
Kevin Readclaudeassisted
2mo ago+7.8K -7
99%Elevated
more lines added than baseline
21
80a196c0feat(gfx10): add MQ3 HFQ3 batched prefill (#298)
Kevin Read
2mo ago+8.6K -274
99%Elevated
more lines added than baseline
22
43dd014efeat(moe+profile): i8 MMQ MoE grouped family + k4 pipeline depth + rocprof coverage
Kaden Schuttclaudeassisted
2mo ago+5.4K -3
99%Elevated
more lines added than baseline
23
b832ba7cMega integration: iterative AWQ/GPTQ plus Kevin PR stack (#296)
Kaden Schutt
2mo ago+9.4K -218
99%Elevated
more lines added than baseline
24
42a83f27Add Python Astrea and Atlas tooling
Kaden Schutt
2mo ago+8.8K -144
99%Elevated
more lines added than baseline
25
b2ee1ba1perf(mq-lloyd): clear ≥120 tok/s ship gate — continuation of #115 (#181)
Kevin Readclaudeassisted
3mo ago+6.3K -92
99%Elevated
more lines added than baseline
26
caa40091feat(gfx906): MMQ kernel redesign + AR decode optimizations — 5× prefill, +16% decode (96% of stock llama.cpp) (#158)
Kevin Readclaudeassisted
3mo ago+7.0K -107
99%Elevated
more lines added than baseline
27
f0f0d9dbtest(quality-gate): refresh gfx1100 MQ4 baselines on dflash (branch-local)
Kaden Schuttclaudeassisted
3mo ago+10.6K -13.5K
99%Elevated
more lines added than baseline
28
438894a1perf(wmma): 4.1× prefill via WMMA GEMM + turbo→givens cleanup
Kaden Schuttclaudeassisted
4mo ago+13.4K -21.2K
99%Elevated
more lines added than baseline
29
1dc18633feat(quality-gate): add 9B MQ4 gfx1010 baselines from BC-250
Kaden Schuttclaudeassisted
4mo ago+6.0K -0
99%Elevated
more lines added than baseline
30
76b17fe5feat(quality-gate): add 4B MQ4 gfx1010 baselines from BC-250
Kaden Schuttclaudeassisted
4mo ago+6.0K -0
99%Elevated
more lines added than baseline
31
31db0ef3feat(magnum): add MagnumQuant + butterfly rotation research crate
Kaden Schuttclaudeassisted
4mo ago+5.7K -0
99%Elevated
more lines added than baseline
32
8b0455dffeat(quality-gate): mandatory MQ4 output regression gate + pre-commit hook
Kaden Schuttclaudeassisted
4mo ago+18.7K -0
99%Elevated
more lines added than baseline
33
9fa33b33feat(installer): add native setup wizard and updates
Kaden Schutt
1w ago+4.5K -1.4K
99%Elevated
more lines added than baseline
34
2a960e19feat: replace TypeScript control plane with Rust
Kaden Schutt
3w ago+11.1K -18.8K
99%Elevated
more lines added than baseline
35
f8fe55d5feat(mfp4-E8): commit runtime infra (E8/Lloyd/P GEMV kernels + dispatch wiring)
Kaden Schuttclaudeassisted
2mo ago+2.6K -4
99%Elevated
more lines added than baseline
36
fd9cfefffold(#415): AWQ on sub-4-bit quant arms onto staging (dual-land)
Kaden Schuttclaudeassisted
2mo ago+3.2K -1.6K
99%Elevated
more lines added than baseline
37
590c25d6fold(lfm2moe+minimax): bring arch_id 10/11 onto feature/dispatch-unification (Ship 2)
Kaden Schuttclaudeassisted
2mo ago+5.7K -2
99%Elevated
more lines added than baseline
38
14697380chore: format changed Rust files
Sadara Kael
2mo ago+3.0K -1.7K
99%Elevated
more lines added than baseline
39
9c659c72fix(vision): satisfy changed-file CI
Kaden Schutt
2mo ago+3.2K -999
99%Elevated
more lines added than baseline
40
de2c5155fix(kernel-cache): version cache key + self-heal invalid device images
Nick Woolmer
2mo ago+3.5K -1.1K
99%Elevated
more lines added than baseline
41
2da44875feat(attn): dots.ocr vision and decode kernels (#369)
Kevin Read
2mo ago+4.0K -64
99%Elevated
more lines added than baseline
42
0db91605feat(rdna-compute): WMMA flash attention for asym4 text prefill
Kevin Read
2mo ago+3.2K -4
99%Elevated
more lines added than baseline
43
adfc9a01test: use q8 for DFlash perf gates
Kaden Schutt
2mo ago+4.4K -38
99%Elevated
more lines added than baseline
44
c5c883cefeat(qwen2): add standalone architecture crate (#297)
Kevin Read
2mo ago+5.6K -9
99%Elevated
more lines added than baseline
45
805d0490feat(prefill): Path 1 + Path 2 MoE prefill grouped-WMMA-GEMM
Kaden Schuttclaudeassisted
2mo ago+2.3K -68
99%Elevated
more lines added than baseline
46
391e8d8cfeat(mq3): _mb4 batch-tile fanout for HFQ3 + MQ3-Lloyd (+77% / +68% gfx1151 9B)
Kevin Readclaudeassisted
3mo ago+2.3K -0
99%Elevated
more lines added than baseline
47
77d61727feat(gfx906): HFQ6/MQ6 Phase A — wave64 + dp4a kernel stack (+41% decode, +248% prefill) (#187)
Kevin Read
3mo ago+2.0K -12
99%Elevated
more lines added than baseline
48
335b3362feat(detect): coherence_probe + hipfire-detect lib — user-facing model behavior debugger (#194)
Kaden Schutt
3mo ago+3.6K -0
99%Elevated
more lines added than baseline
49
34579d74MQ3-Lloyd WMMA prefill (#116 Phase 5) — Phases A+B1+B2+B3+C (#195)
Kevin Readclaudeassisted
3mo ago+3.6K -114
99%Elevated
more lines added than baseline
50
df000b69PFlash speculative prefill: Phase 5 ship + regression gate (closes #93) (#135)
Kaden Schuttclaudeassisted
3mo ago+7.1K -28
99%Elevated
more lines added than baseline
  • 37b36c8brefactor(loading): collapse qwen35 to one public load_weights entry
    Author
    Björn Bösel
    When
    1mo ago
    Lines
    +17.8K -5.8K
    Risk
    100%Elevated
  • 09800fbfrefactor(rdna-compute): decompose GPU dispatch object (#342)
    Author
    Björn Bösel
    When
    2mo ago
    Lines
    +35.2K -34.2K
    Risk
    100%Elevated
  • a18d694achore(quality-eval): generate eval slice + populate canary fixture
    Author
    Kevin Read
    When
    3mo ago
    Lines
    +45.8K -31
    Risk
    100%Elevated
  • a020c8e1feat(cdna): MI300X calibration optimization (Phase 2-5) + gfx942 baseline + wave64 residual gemv (#90)
    Author
    Kaden Schutt
    When
    3mo ago
    Lines
    +65.4K -48
    Risk
    100%Elevated
  • bccc9a17wmma correctness fix + MQ6 family + cross-arch prefill + gate framework
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +35.4K -11.8K
    Risk
    100%Elevated
  • 72620fc9Track C: HFQ4 KV cache — 5 experiments completed
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +72.3K -56
    Risk
    100%Elevated
  • a9dc252fPer-layer profiling infrastructure + full benchmark report
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +482.6K -0
    Risk
    100%Elevated
  • 94d0bf2bfeat(runtime): FireMap - grow Qwen3.5 KV on demand with HIP VMM (#549)
    Author
    黃思銘-HKUST
    When
    1w ago
    Lines
    +8.1K -746
    Risk
    99%Elevated
  • a064932cfeat(kernels): lower radiowave into beta and compile kernels through it
    Author
    Kaden Schutt
    When
    2w ago
    Lines
    +9.0K -20
    Risk
    99%Elevated
  • 64f16d22docs(skills): promote canonical workflow overhaul
    Author
    Kaden Schutt
    When
    3w ago
    Lines
    +4.0K -3.2K
    Risk
    99%Elevated
  • 1a0acd11docs: promote lfm-redline documentation foundations
    Author
    Kaden Schutt
    When
    3w ago
    Lines
    +8.5K -4.9K
    Risk
    99%Elevated
  • 3240ad0edispatch: graft redline replay behind fail-closed gate
    Author
    Kaden Schutt
    When
    4w ago
    Lines
    +12.7K -0
    Risk
    99%Elevated
  • 53bdf083feat: agentic PR merge-gate (Tier-3 GPU gate + autoresearch/ar package)
    Author
    Kaden Schutt
    When
    4w ago
    Lines
    +11.3K -0
    Risk
    99%Elevated
  • 1876bd8efix(gate-up): make nosync variants buildable and opt-in
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +5.8K -2.3K
    Risk
    99%Elevated
  • 0be4ad16refactor(paro): remove stale env-gated experiments
    Author
    Björn Bösel
    When
    2mo ago
    Lines
    +15.0K -7.3K
    Risk
    99%Elevated
  • 0884535cQwen3.5/3.6 grammar-guided tool calling and prompt caching
    Author
    Nick Woolmer
    When
    2mo ago
    Lines
    +6.5K -234
    Risk
    99%Elevated
  • 1cafa192[codex] Land A3B MTP and DFlash perfmaxx work (#352)
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +5.6K -1.5K
    Risk
    99%Elevated
  • 15a96b98research(paroquant): add G256 milestone probes
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +6.5K -676
    Risk
    99%Elevated
  • b263fb3cfeat(arch-deepseek4): add DeepSeek V4 Flash support (arch_id=9)
    Author
    Nick Woolmer
    When
    2mo ago
    Lines
    +25.3K -3.1K
    Risk
    99%Elevated
  • 6c014fc6dots-ocr: phase 0 item 5 — HF + vLLM reference captures
    Author
    Kevin Read
    When
    2mo ago
    Lines
    +7.8K -7
    Risk
    99%Elevated
  • 80a196c0feat(gfx10): add MQ3 HFQ3 batched prefill (#298)
    Author
    Kevin Read
    When
    2mo ago
    Lines
    +8.6K -274
    Risk
    99%Elevated
  • 43dd014efeat(moe+profile): i8 MMQ MoE grouped family + k4 pipeline depth + rocprof coverage
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +5.4K -3
    Risk
    99%Elevated
  • b832ba7cMega integration: iterative AWQ/GPTQ plus Kevin PR stack (#296)
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +9.4K -218
    Risk
    99%Elevated
  • 42a83f27Add Python Astrea and Atlas tooling
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +8.8K -144
    Risk
    99%Elevated
  • b2ee1ba1perf(mq-lloyd): clear ≥120 tok/s ship gate — continuation of #115 (#181)
    Author
    Kevin Read
    When
    3mo ago
    Lines
    +6.3K -92
    Risk
    99%Elevated
  • caa40091feat(gfx906): MMQ kernel redesign + AR decode optimizations — 5× prefill, +16% decode (96% of stock llama.cpp) (#158)
    Author
    Kevin Read
    When
    3mo ago
    Lines
    +7.0K -107
    Risk
    99%Elevated
  • f0f0d9dbtest(quality-gate): refresh gfx1100 MQ4 baselines on dflash (branch-local)
    Author
    Kaden Schutt
    When
    3mo ago
    Lines
    +10.6K -13.5K
    Risk
    99%Elevated
  • 438894a1perf(wmma): 4.1× prefill via WMMA GEMM + turbo→givens cleanup
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +13.4K -21.2K
    Risk
    99%Elevated
  • 1dc18633feat(quality-gate): add 9B MQ4 gfx1010 baselines from BC-250
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +6.0K -0
    Risk
    99%Elevated
  • 76b17fe5feat(quality-gate): add 4B MQ4 gfx1010 baselines from BC-250
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +6.0K -0
    Risk
    99%Elevated
  • 31db0ef3feat(magnum): add MagnumQuant + butterfly rotation research crate
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +5.7K -0
    Risk
    99%Elevated
  • 8b0455dffeat(quality-gate): mandatory MQ4 output regression gate + pre-commit hook
    Author
    Kaden Schutt
    When
    4mo ago
    Lines
    +18.7K -0
    Risk
    99%Elevated
  • 9fa33b33feat(installer): add native setup wizard and updates
    Author
    Kaden Schutt
    When
    1w ago
    Lines
    +4.5K -1.4K
    Risk
    99%Elevated
  • 2a960e19feat: replace TypeScript control plane with Rust
    Author
    Kaden Schutt
    When
    3w ago
    Lines
    +11.1K -18.8K
    Risk
    99%Elevated
  • f8fe55d5feat(mfp4-E8): commit runtime infra (E8/Lloyd/P GEMV kernels + dispatch wiring)
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +2.6K -4
    Risk
    99%Elevated
  • fd9cfefffold(#415): AWQ on sub-4-bit quant arms onto staging (dual-land)
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +3.2K -1.6K
    Risk
    99%Elevated
  • 590c25d6fold(lfm2moe+minimax): bring arch_id 10/11 onto feature/dispatch-unification (Ship 2)
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +5.7K -2
    Risk
    99%Elevated
  • 14697380chore: format changed Rust files
    Author
    Sadara Kael
    When
    2mo ago
    Lines
    +3.0K -1.7K
    Risk
    99%Elevated
  • 9c659c72fix(vision): satisfy changed-file CI
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +3.2K -999
    Risk
    99%Elevated
  • de2c5155fix(kernel-cache): version cache key + self-heal invalid device images
    Author
    Nick Woolmer
    When
    2mo ago
    Lines
    +3.5K -1.1K
    Risk
    99%Elevated
  • 2da44875feat(attn): dots.ocr vision and decode kernels (#369)
    Author
    Kevin Read
    When
    2mo ago
    Lines
    +4.0K -64
    Risk
    99%Elevated
  • 0db91605feat(rdna-compute): WMMA flash attention for asym4 text prefill
    Author
    Kevin Read
    When
    2mo ago
    Lines
    +3.2K -4
    Risk
    99%Elevated
  • adfc9a01test: use q8 for DFlash perf gates
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +4.4K -38
    Risk
    99%Elevated
  • c5c883cefeat(qwen2): add standalone architecture crate (#297)
    Author
    Kevin Read
    When
    2mo ago
    Lines
    +5.6K -9
    Risk
    99%Elevated
  • 805d0490feat(prefill): Path 1 + Path 2 MoE prefill grouped-WMMA-GEMM
    Author
    Kaden Schutt
    When
    2mo ago
    Lines
    +2.3K -68
    Risk
    99%Elevated
  • 391e8d8cfeat(mq3): _mb4 batch-tile fanout for HFQ3 + MQ3-Lloyd (+77% / +68% gfx1151 9B)
    Author
    Kevin Read
    When
    3mo ago
    Lines
    +2.3K -0
    Risk
    99%Elevated
  • 77d61727feat(gfx906): HFQ6/MQ6 Phase A — wave64 + dp4a kernel stack (+41% decode, +248% prefill) (#187)
    Author
    Kevin Read
    When
    3mo ago
    Lines
    +2.0K -12
    Risk
    99%Elevated
  • 335b3362feat(detect): coherence_probe + hipfire-detect lib — user-facing model behavior debugger (#194)
    Author
    Kaden Schutt
    When
    3mo ago
    Lines
    +3.6K -0
    Risk
    99%Elevated
  • 34579d74MQ3-Lloyd WMMA prefill (#116 Phase 5) — Phases A+B1+B2+B3+C (#195)
    Author
    Kevin Read
    When
    3mo ago
    Lines
    +3.6K -114
    Risk
    99%Elevated
  • df000b69PFlash speculative prefill: Phase 5 ship + regression gate (closes #93) (#135)
    Author
    Kaden Schutt
    When
    3mo ago
    Lines
    +7.1K -28
    Risk
    99%Elevated
Showing 50 of 3,072 commits

How the score behaves here

Change risk →

Two views of the same model: where the cuts fall, and what commit shape lands you above them.

Score distribution

Every scored commit, binned on the raw 0 to 10 score rather than the percentile. Percentile ranks are uniform by construction, so that axis has no shape to draw. The dashed lines are the tercile cuts behind each row's priority pill.

0409typical ↑↑ elevated0.05.010.0Change-risk score →
Below typical
Typical
Elevated

Size against diffusion

The 200 most recent commits, on their own recency sample rather than the feed above: that defaults to risk-sorted, so reusing it would plot only the top tercile and call it the spread. Big and scattered is what the model penalises. Click a dot to open it.

1101001,00010,000Lines changed (log) →0.04.9Diffusion →
Below typical72
Typical54
Elevated74

Commit history for rakaarwaky/hipfire

rakaarwaky/hipfire has 3,279 commits in its history from 37 contributors, the first of them Mar 20, 2026. In the last 90 days 921 files were touched, 5,186 times in total, most often kernels.rs. Every commit is scored for change risk from its size, spread and the history of the files it touches.