Niko1221/Strata

0 Fork 0

mirror of https://github.com/Niko1221/Strata.git synced 2026-11-10 11:51:31 -10:00

Files

T

History

Niko1221 3d26434f21 tools/: 27 untested, undocumented experiment scripts move to research/tools (3 folders, a README each saying what and when); two machine-local DFlash2 analysis scripts move to history/tools; documented, tested and packaged tools stay; tools/dflash2 files a training branch changes are left

2026-11-07 23:37:51 +01:00

..

check_metal_dump.py

metal: DFlash2 Q8_0 GEMM kernel over a repacked pack (df2_mm: 12.6 ms a round against 36), Q8_0 draft head, STRATA_DF2_DUMP and tools/dflash2/check_metal_dump.py (final rows cos 0.9999 against the torch port)

2026-11-01 23:12:16 +01:00

dflash2_ref.py

Remove unused imports reported by Ruff F401 (cleanup 4.3)

2026-11-07 16:28:41 +01:00

mk_parity_case.py

DFlash2 drafter for Qwen3.8-Flash-Next: Q8_0 pack tool, GPU forward (conv, K/V ring, windowed attention, top-16 selector), permanent verify/prefill taps, --draft dflash2 --draft-model (opt-in, MTP stays the fallback)

2026-10-21 14:04:15 +02:00

mk_prompts.py

tools/dflash2: the torch port of the DFlash2 drafter and the scripts of the acceptance experiment (research tooling)

2026-10-21 13:24:21 +02:00

mk_serve_scenario.py

wip: DFlash2 in the serve loop (ring stamps, request history, prompt rows, spec loop), serve test harness

2026-10-31 19:04:57 +01:00

orch.py

tools/dflash2: the torch port of the DFlash2 drafter and the scripts of the acceptance experiment (research tooling)

2026-10-21 13:24:21 +02:00

q8test.py

Remove unused imports reported by Ruff F401 (cleanup 4.3)

2026-11-07 16:28:41 +01:00

README.md

tools/: 27 untested, undocumented experiment scripts move to research/tools (3 folders, a README each saying what and when); two machine-local DFlash2 analysis scripts move to history/tools; documented, tested and packaged tools stay; tools/dflash2 files a training branch changes are left

2026-11-07 23:37:51 +01:00

run_all2.py

tools/dflash2: the torch port of the DFlash2 drafter and the scripts of the acceptance experiment (research tooling)

2026-10-21 13:24:21 +02:00

run_eval.py

tools/dflash2: the torch port of the DFlash2 drafter and the scripts of the acceptance experiment (research tooling)

2026-10-21 13:24:21 +02:00

run_one.ps1

tools/dflash2: the torch port of the DFlash2 drafter and the scripts of the acceptance experiment (research tooling)

2026-10-21 13:24:21 +02:00

serve_test.py

wip: DFlash2 in the serve loop (ring stamps, request history, prompt rows, spec loop), serve test harness

2026-10-31 19:04:57 +01:00

test_port.py

Tests: the not-preallocated check is the full size; tools/dflash2/test_port.py skips without its lab files

2026-11-07 16:12:41 +01:00

README.md

DFlash2 acceptance experiment

Research tooling, not shipped with the app. It measures how many tokens per round the incoai DFlash2 drafter for Qwen3.8-Flash-Next would accept on Strata's quantized target. The measurement is offline, from hidden states the engine dumps. It is for developers who want to repeat or extend the experiment.

The scripts keep the paths of the original run (C:\dflash2-work on the development PC, C:\strata-test\dflash2 on the test box). Edit the paths before you reuse them. The method and results are in the project's internal notes and are not part of this tree.

Engine variables

Both are off unless set.

Variable	Effect
`STRATA_TAP_FILE=<file>`	Writes `pos, token, taps` records. `Verifier::record_window` copies each verified row's completed layer output (the mean of the 4 hyper-connection streams, the quantity the drafter was trained on) and `Verifier::run` appends it.
`STRATA_TAP_LAYERS`	Layers to tap, default `5,15,25,35,45`. Used with `STRATA_TAP_FILE`.
`STRATA_SPEC_T_CYCLE=1,2,...,8`	CLI bench only. The verify window size cycles; the report line `window cost (ms/round)` gives the cost of each width.

Scripts

Script	Purpose
`mk_prompts.py`	Builds chat-template prompt ids with the server's own tokenizer and template.
`orch.py <prompt> <max_new> [gen,genf,genfm,tap]`	Drives `run_one.ps1` on the test box: free generation (ids and MTP statistics), the same text with MTP only, and the teacher-forced tap dump with oracle drafts (`--spec 8 --spec-oracle --spec-follow`, a 2-token prompt, because a native pack never starts the speculative loop from a 1-token prompt).
`dflash2_ref.py`	Plain-torch port of z-lab's `DFlash2DraftModel`, with the target's embedding and head dequantized from the GGUF. `test_port.py` checks it against the original.
`run_eval.py`	Walks the tap file with a few block sizes.
`run_all2.py`	Evaluates every position, resumable: per-anchor correctness and the selector's confidences.
`analyze2.py`, `final_tables.py`	Moved to history/tools/dflash2: policies, and speed estimates from the measured window costs. They read result files that only exist in `C:\dflash2-work` on the original PC.