F1-F3: NextN geometry on colibri line + F2 current-turn hits + F3 bit-length heat - #157
Merged
Merged
Conversation
…ifact routes
Stage-C/D owner ruling: the fork ships the two atlas files; atlas-web proxies
GET {engine}/experts.json and propagates the upstream status verbatim
(server.ts:250-320), so a missing/refused artifact must surface as 404 —
never a misleading 200, never another model's atlas.
- server-atlas: atlas_file()/atlas_json() — resolution HYDRA_EXPERT_ATLAS
(dir override) -> model file's own directory; engine-id refusal gate on
provenance.engine_id (exact 16-hex FNV geometry id or the
"$arch:$basename[:$size]" short form); 0/404/500 contract (404 missing or
refused, 500 corrupt/unreadable). Geometry gains model_path/total_size;
load_impl scratch-state reset on partial init keeps dense-model results
fully cleared (byte-identity argued; single-shard MoE re-verified live).
- server.cpp: routes /experts.json + /v1/experts.json (observability tier)
and /expert-ranks.json + /v1/expert-ranks.json (ranking tier twins).
- tests/test-hydra-atlas-file: synthetic 2-row MoE GGUF; 22 checks covering
happy paths (both id spellings), 404 missing/wrong-id/no-provenance,
500 corrupt JSON, tier independence, unknown kind.
🤖 Generated with Codebuff
Co-Authored-By: Codebuff <noreply@codebuff.com>
With NextN models (qwen35moe, nextn_predict_layers=1), llama-model counts
the nextn layer inside block_count: n_layer_all=41, trunk n_layer()=40.
The MTP tensors live in blk.40 (blk.40.nextn.* alongside blk.40.ffn_*_exps),
so the old geometry init (trunk = 0..block_count-1) classified blk.40 as a
trunk MoE row. The topk hook then reported it dead forever ("atlas: 1/41
MoE rows lack topk tensors (dead layers: 40)") and the EMAP row read as a
permanently unrouted VRAM tier row — misleading on the atlas UI.
Fix: trunk = block_count - nextn; blocks >= trunk with exps tensors are
classified as nextn_rows. Geometry now reports moe_rows 0..39 and
nextn_rows [40]; consumers (UI /turns labels, EMAP grid) can render the
MTP draft layer honestly.
Found via track t-34d7bdf5d6 (owner spotted layer 40 never routing).
🤖 Generated with Codebuff
Co-Authored-By: Codebuff <noreply@codebuff.com>
…m parity)
F2 (architect package d-9981fa1092): the per-step hits bitmap never
zeroed between decode steps, so g_hits_step OR-accumulated lifetime bits
and /experts served a sticky cumulative hits picture. Add
hydra_atlas::begin_step() (zeroes g_hits_step under lock), call it from
the decode hook before the per-row accumulate loop (server-context),
serve /experts hits from the current TURN window (g_turn_hits, folded in
end_step, cleared at record_turn) with a size-guard, and clear the step
bitmap in end_step as belt-and-braces. NO clear-on-read semantics.
F3: EMAP heat encoding switches from linear-clamp-63 to bit-length,
verbatim port of upstream c/telemetry.h emap_emit
(`while(u){heat++;u>>=1;} if(heat>63) heat=63;`). Precondition checked:
sweep_analyze_790.py and colibri_gates.py never decode EMAP bytes
(heat from raw probe counts); sweep.sh deltas a monotone byte (encoding-
agnostic; saturates-at-63 comment still true).
Tests: test-hydra-atlas-file extended (+10 checks) — window hits = step
bits, OR within turn, record_turn clears served window (sticky-bug
regression), per-row popcount <= k*n_out, and heat = bit-length at
counts 1/3/4/70 (emap_emit parity). CPU build: test target compiles
clean; ./test-hydra-atlas-file: 31/31 ok, exit 0.
🤖 Generated with Codebuff
Co-Authored-By: Codebuff <noreply@codebuff.com>
…as) + N1 trunk guard Architect re-review d-7045376b02 on PR #157: B1 (blocking, script-only): sweep.sh computed probe selections as after-before deltas on EMAP bytes (map & 63). With F3 bit-length heat those deltas are not selection counts — hot experts saturate (127 stays heat 7, delta 0, drops out) while cold experts dominate, inverting the affinity/lift ranking downstream (analyze.py sums, validate.py shares treat "selections" as linear counts). The sweep now reads the probe turn's exact routing counts from the newest /turns/<seq> record (same decode hook as the counters; {row, expert, count} mapped through geometry.moe_rows/nextn_rows), records source:"turns-routing" + turn_seq for provenance, and falls back to telemetry:false when the ring did not advance. The before.json snapshot is kept only for its seq and telemetry_enabled gate. N1 (nit): load_impl returns false cleanly on negative trunk (malformed-GGUF insurance; llama-model asserts n_layer_nextn <= n_layer_all before the atlas loads, but the atlas parses independently). Verification: bash -n sweep.sh OK; test-hydra-atlas-file rebuilt clean, 31/31 "all ok" exit 0. PR description corrected: the F3 commit's "sweep.sh deltas a monotone byte (encoding-agnostic)" claim was WRONG (consumers treat deltas as linear counts) — superseded by this fix. 🤖 Generated with Codebuff Co-Authored-By: Codebuff <noreply@codebuff.com>
…d 2) Architect round-2 review on PR #157 (e8f271b): B2 (blocking): the stats-writer argv indexes were off by one — argv[3] got the category, int(argv[4]) crashed on it ("invalid literal for int(): 'math'"), and the output opened argv[2] (a file named after the turn seq). set -uo pipefail without -e made every probe fail silently. Now: category argv[4], idx int(argv[5]), output open(argv[3]). B3 (blocking): BEFORE_SEQ read /experts .seq = g_seq (decode-step counter), while TURN_SEQ is g_turn_seq (turn counter) — the comparison was always false after any decoding, so every probe after the first would record telemetry:false. Now the before-value comes from GET /turns .seq, snapshotted BEFORE generation; /experts is polled only for the telemetry_enabled gate. Should-fixes: exact-one-turn gate (TURN_SEQ == BEFORE+1) — concurrent client turns no longer silently counted as the probe's selections; mismatch writes telemetry:false with reason (no-new-turn | ring-advanced-by-N | telemetry-disabled | turns-unavailable). Row filter corrected to row < len(moe_rows) (comment now matches code). Hermetic smoke (architect-required, stub engine with /experts.seq ~900K >> /turns.seq ~50 to exercise B3): 4 probes → 4 sidecars telemetry:true, correct category/idx, source:turns-routing, linear counts, MTP row excluded, turn_seq advancing; analyze.py exits 0 (3 experts, 2 categories). bash -n clean. 🤖 Generated with Codebuff Co-Authored-By: Codebuff <noreply@codebuff.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Architect handoff package (hydra_vortex track t-34d7bdf5d6, decision d-9981fa1092), delivered on the rebased feat/flash-next-colibri line (base 6b7dc6a).
Commits
0abcf9287+3840b2de3are already patch-equivalent on the base (verified via git cherry); only the artifact-routes commit was missing.nextn_rows, not a trunk row (re-applied 6dbf575). Merged hunk verified semantically: vector sizedblock_count + nextn,trunk = block_count - nextn, moe_rows 0..trunk-1, nextn_rows from trunk. EANmtp_excluded:truealready present on this line.F2 — sticky hits fix
hydra_atlas::begin_step()zeroesg_hits_stepunder lock; decode hook (server-context.cpp) calls it once per step BEFORE the per-row accumulate loop./expertsnow serves the CURRENT TURN window (g_turn_hits, folded per step inend_step, cleared atrecord_turn) — never the per-step bitmap. NO clear-on-read.end_stepalso clears the step bitmap (belt-and-braces: a skippedbegin_stepcan never resurrect stale bits).F3 — heat encoding: bit-length (upstream parity)
encode_cell: verbatim port of upstreamc/telemetry.h emap_emit:while(u){heat++;u>>=1;} if(heat>63) heat=63;(was linear-clamp-63).sweep_analyze_790.pyandcolibri_gates.pynever decode EMAP bytes (heat computed from raw probe counts);tools/expert-atlas/sweep.shcomputes deltas over the monotone byte (encoding-agnostic; the saturates-at-63 note remains true).Verification (local, CPU-only)
-DGGML_CUDA=OFF -DLLAMA_CURL=OFF, 20-core -j16.test-hydra-atlas-filebuilds clean; 31/31 checks ok, exit 0 — including 10 new F2/F3 checks: window hits = current step bits; OR-accumulation within a turn;record_turnclears the served window (sticky-bug regression); per-row popcount <= k*n_out; heat = bit-length at counts 1/3/4/70 (emap_emit parity; pre-F3 linear would say 1,3,4,63).Not in this PR
Correction (architect re-review d-7045376b02, fixed in e8f271b)
The F3 commit message's claim that sweep.sh deltas are "a monotone byte (encoding-agnostic)" was wrong:
analyze.pysums andvalidate.pyshares treat the recorded "selections" as linear counts, so EMAP-byte deltas under bit-length heat are not counts (hot experts saturate → delta 0 → vanish; cold dominate) and would invert the affinity ranking. Superseded: sweep.sh now reads the probe turn's exact per-turn routing counts fromGET /turns/<seq>(source: "turns-routing",turn_seqprovenance, MTP rows excluded) and never decodes EMAP bytes. Also includes N1: cleanreturn falseon negative trunk inload_impl. Verified:bash -nOK,test-hydra-atlas-file31/31 exit 0.