fix(atlas): EAN ggml_sqr contiguity + POST /atlas/reset?scope=ean|all (hydra_vortex#806) - #158
Merged
ddvnguyen merged 1 commit intoSep 27, 2026
Conversation
… (hydra_vortex#806)
RULING 1 (docs/design-moe-demand-admission-phase15.md:1850-1854): the EAN
per-slot reduction ran ggml_sqr on a strided ggml_view_2d, tripping
ggml_cuda_op_sqr's contiguity assert at load warmup whenever
HYDRA_EAN_STATS=1 (reproduced pre-fix: unary.cu:144 abort in 6.7 s).
Materialize one contiguous copy (ggml_is_contiguous ? slot : ggml_cont)
plus a build-time GGML_ASSERT before the sqr; the whole change stays
inside the getenv gate, so env-unset graphs stay byte-identical
(OFF-parity structural). The conditional means decode graphs
(n_tokens==1, view passes ggml_is_contiguous) add zero ops.
RULING 3 (:1864-1868): new POST /atlas/reset?scope=ean|all backed by
hydra_atlas::reset_scope() — scope=ean zeroes g_ean_gxn/g_ean_nsel in
place under g_mtx (sizes kept, /experts shows ean cells=0 until the next
decode refills), scope=all delegates to the existing reset();
deliberately separate from the zero-caller per-probe reset(). 400 on an
unknown scope.
Also reword the stale in-tree pinfile-parser citations in
tools/expert-atlas/{README.md,export_pinfile.py} (hydra_cpu_init /
ggml-cuda.cu parsers are not present in this tree; format contract only).
Tests: A/B smoke on RTX 5060 Ti, model Ornith-1.5-35B-A3B Compact,
--n-cpu-moe 12: pre-fix binary aborts at warmup in 6.7 s (unary.cu:144,
the exact defect); post-fix with HYDRA_EAN_STATS=1 HYDRA_EXPERT_META=1
reaches listening in 39 s, /health 200, completion 200, /experts ean
accumulates 1185 cells, POST scope=ean 200 (cells 0 -> 702 refill after
the next completion), scope=all 200, bad scope 400. Env-unset run:
normal start, byte-identical greedy completion, ean key absent.
build-colibri-parity llama-server builds clean.
Refs ddvnguyen/hydra_vortex#806
Co-Authored-By: Oh My Pi (opencode-go/mimo-v2.6-flash)
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.
What
Closes ddvnguyen/hydra_vortex#806 (RULING 1 + RULING 3,
docs/design-moe-demand-admission-phase15.md:1850-1868).src/llama-graph.cpp, inside theHYDRA_EAN_STATSgate): the per-slot reduction ranggml_sqron a stridedggml_view_2d→ggml_cuda_op_sqrcontiguity assert aborted warmup. Nowslot_c = ggml_is_contiguous(slot) ? slot : ggml_cont(ctx0, slot)+ build-timeGGML_ASSERT. Conditional single copy: decode graphs (n_tokens==1) add zero ops; env-unset graphs are byte-identical (OFF-parity structural).POST /atlas/reset?scope=ean|all(tools/server/server.cpp+server-atlas.{cpp,h}): newhydra_atlas::reset_scope();scope=eanzeroesg_ean_gxn/g_ean_nselin place under lock (sizes kept →/expertsshowseancells=0 until the next decode refills),scope=alldelegates toreset(), 400 on unknown scope. Deliberately separate from the zero-caller per-probereset().tools/expert-atlas/{README.md,export_pinfile.py}reworded (parser absent; format contract only).Evidence (A/B smoke, RTX 5060 Ti, Ornith-1.5-35B-A3B Compact,
--n-cpu-moe 12)HYDRA_EAN_STATS=1unary.cu:144 GGML_ASSERT(ggml_is_contiguous(src0))— exact ggml-org#806 defectHYDRA_EAN_STATS=1 HYDRA_EXPERT_META=1/health200, completion 200,/expertseanaccumulates 1185 cells,POST scope=ean200 (cells 0 → 702 refill next completion),scope=all200, bad scope 400eankey absentBuild:
build-colibri-parityllama-server clean (CUDA 13.2.1, sm_120).Refs ddvnguyen/hydra_vortex#806