Cold prefill
256,000 input tokens · tok/s ↑A complete input read from an empty cache. Halogen 0.17.1 was not measured on this panel.
STRIX HALO / THE OCTOBER 12 RECORD
Tom Turney
Expert Optimization
saves ROCmFPX.
The same HC-Q8 weights. A rebuilt ROCmFPX stack.
Follow the old build and Tom’s new build through the race, alongside Ciru IU4, Gufo, and the newest Halogen run.
TOM BUILD / NATIVE PP81921,236.47tokens / second
Official ROCm 7.2.18192 batch · 8192 microbatchAll 8 fast switches · MTP3 serving
99 / 100 Hermes19 passes · one preserved failureOnly available matching pass
Standalone native prefill.
Warmup excluded. One scored round.
Warmup excluded. One scored round.
THE TOM TURNEY EDITION / RECORDED REPLAY
The Tool Gauntlet.
Five stacks. Fifteen checkpoints. Every outcome stays on the track.
SCORED TASK WALL CLOCK00:00.0010× playback
SPEED
START00:00
PassPartialFail— Not measuredHover or select a checkpoint for evidence.
Race feed
Live scores · saved runs
Source order · separate historical protocols
Checkpoint inspector · select any track light
CHECKPOINT INSPECTOR
Every light has a story.
Select a checkpoint for its verdict, measured duration, executed tools, and verifier explanation.
01 / READ. GENERATE. COMPARE.
Where the new build moves.
Recorded observations, with the original build on its own row. Higher rates mean faster processing.
NOT A CONTROLLED RERUN
Decode
HE0–9 · (n−1) / decode time · tok/s ↑Thinking off. Code generation speed, with correctness unscored. Halogen’s stock tool-workload rate is shown separately below.
The cold-prefill sweep
32K → 256K tokens · empty initial cacheThe older and October 12 records use different CPU governors, serving batches, runtime revisions and workload history. These plots compare saved stack observations; they do not isolate Tom’s changes as a controlled causal speed gain.
Explore append prefill at cached depths
THE NEWEST HALOGEN RECORD / OCT 07–08
0.17.1. Stock server.
Five preserved passes.
This is the newest saved Halogen run, using its native frontend. The race replays the best hard-tools pass: pass 2, 23/30 points. Pass 4 ties its score; pass 2 has the shorter task wall. The table keeps all five passes visible.
These PP/TG rates come from the scored 84-case tool workload with retained cache. They belong to this workload and are separate from cold prefill and HE0–9.
02 / THE FIVE-STACK SCORECARD
Keep the tradeoffs in view.
Dates and panel coverage stay with every result. Missing measurements are left blank.
Each race circuit and its scorecard use the best whole saved pass for that stack: highest circuit score, with shortest recorded task wall breaking ties. Pass numbers and available counts are shown; tasks and their timings stay together in the selected pass. Halogen’s newest stock protocol has no new Hermes, HE0–9, fidelity or cold-prefill result.
Inspect numerical fidelity and deployment sizes
Teacher KL divergence
2,048 full-vocabulary positions · lower is closerWeight / deployment package
GiB · saved package accountingPackage coverage differs by stack; HC-Q8 is the GGUF weight file. This is disk footprint, not separate discrete VRAM.
03 / GIVE CREDIT. KEEP THE RECEIPTS.
The stack behind the headline.
Tom’s real runtime
Unmodified ROCmFPX c4aa38b, including merged PRs 37–52. Official AMD ROCm 7.2.1 libraries were verified in the running processes. Serving uses batch and microbatch 8192, 16 threads, f16 KV, flash attention and MTP3 plus ngram.
The same HC-Q8 weights
The new and original HC-Q8 records use the pinned candidate weights. All eight fast switches are enabled. Model-mode sampling inherits the pinned Qwen base card. Fidelity uses target-only logits; native PP8192 generates no output tokens.
Pinned HC-Q8 model card ↗A replay of measured work
The animation interpolates each task’s measured duration and reveals its recorded verdict at completion. Individual tool calls have no playback timestamps. No model inference was rerun to create this page.
The original race dashboard ↗Best-pass selection, protocol differences and preserved failures
These saved records are not a single controlled head-to-head trial. Earlier Ciru IU4, Gufo and original HC-Q8 results come from the original comparison; Tom’s build is the October 12 qualification. Host policies, batches, run dates, warm-history and lifecycle differ.
Halogen 0.17.1 uses the stock 84-case aligned-tool-rules protocol; its race lane shows TC70–84 from its best hard-tools pass (pass 2 of 5). This differs from the standalone corrected 15-case panel used for Tom’s build. The original scorers are preserved, including known TC34/TC40/TC80 caveats. No overall leaderboard or combined winner is inferred.
Tom’s Hermes run retains HA-17’s score of 70: delegated tools executed, but the merged summary failed its artifact check. Counting retains 17/18 clean first outputs; the 4096-input, empty-depth output skipped 75 and 76. A two-request fixture-mount failure is preserved as invalid harness diagnostics and excluded from the matching Hermes pass; it was corrected before that full 20-case pass.
The current native PP8192 result has one scored round after the required unscored native warmup. Historical Hermes records have two runs: the race selects Ciru IU4 pass 2, Gufo pass 2, and original HC-Q8 pass 1 by score. Tom has one matching full pass. Every selected whole pass retains its original outcomes and task timings. No cases are combined across passes, and no model inference was rerun for this selection.
This is a Ciru research page honoring Tom’s work. The theme draws from his public profile, handle and the supplied portrait.