For streamed (non-resident) experts the MoE forward is disk-bound: each
64-expert block first blocks on a parallel pread of every miss, then runs
the matmuls serially. Those two phases don't overlap, so the compute
cores idle while the block's weights are still coming off NVMe.
PIPE=1 hands the misses' expert_load() to a small persistent pool of I/O
worker pthreads and lets the main thread start matmul immediately. The
main thread walks the block's experts in routing order and waits on a
per-slot ready flag only for the expert it needs right now, so loads of
later experts in the block hide behind matmul of earlier ones. All
matmul_qt stays on the main thread (it parallelises internally via
OpenMP and gates GPU dispatch on !omp_in_parallel()), so the I/O pool
never competes with the compute team for the matmul itself.
Cross-generation correctness: generation-tagged lock-free cursor
--------------------------------------------------------------------
The batch state (njobs/eids[]/layer/ready[]) is reused in place across
64-blocks, so a straggler or late-woken worker could touch the NEXT
batch's state. Two prior fixes (a drain barrier, then an _Atomic active
counter) each still left a cross-generation window. Both are now removed
and replaced with a single generation-tagged cursor:
_Atomic uint64_t cur = (gen << 8) | index; // gen main-only, index 0..njobs
- dispatch writes njobs/layer/eids[]/ready-reset with RELAXED stores,
then RELEASE-stores cur (bumping gen); that release publishes all
batch state to any worker whose ACQUIRE-load of cur sees the new gen.
- a worker grabs a job by CAS-advancing the index; it reads eids[i]/
layer only AFTER the winning CAS. The CAS comparand carries the
generation, so if a new batch was published the stale CAS fails and
the worker re-reads — it can never grab a wrong-generation job or
read torn state, no matter where it was preempted (wake gap,
post-cursor, anywhere).
- gen is bumped only by the main thread and is monotonic ⇒ no ABA.
- the per-expert pipe_wait(ready[q]) in the matmul loop (kept) makes
every grabbed job complete before the block ends, so no grab
outlives its generation. That is what makes the old `active`
counter and the end-of-block drain barrier unnecessary — both are
removed. ready[] is reset before the publishing release, so no stale
flag survives into the new generation.
- the mutex/condvar now exist ONLY to park and wake idle workers, not
for correctness.
This only reorders I/O, never the computation: greedy decode output is
byte-identical to the blocking path.
Default OFF (PIPE=0) so upstream behaviour is unchanged; PIPE=1 opts in
and PIPE_WORKERS=n sizes the pool (default 8). No Makefile change: pure C
(stdatomic.h + sched.h), builds with the default toolchain.
Claude-Session: https://claude.ai/code/session_01DS7oc65c5RdA9V99otRCwt
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The MoE forward does many tiny, back-to-back per-expert matmul regions.
Under libgomp's default passive wait policy the worker team is parked
between regions, and the thread re-wake latency dominates the actual
compute. Switching the team to an active wait policy (with a bounded spin
count so long NVMe expert-load stalls still yield instead of burning a
core) keeps the threads hot across those regions.
Measured on the Zen5 build: expert-matmul wall time 66.9s -> 20.9s
(~3.2x on that phase). Output is numerically unchanged — this only
affects thread scheduling, not the computation.
Mechanism: libgomp parses OMP_/GOMP_ vars in a constructor that runs
before main(), so setenv() from main() and continuing is too late
(verified: the already-initialised runtime ignores it). Instead we seed
the winning defaults on first entry with overwrite=0 (so any value the
user already set wins), set a COLI_OMP_TUNED sentinel, and execv() self
once so a fresh libgomp constructor reads them. The sentinel guards the
exec, so we re-exec at most once. The block is the first statement in
main() so argv is passed verbatim to execv().
Discoverability / observability (review follow-up):
- COLI_NO_OMP_TUNE=1 is a documented kill-switch that disables the
whole re-exec + tuning path in one shot (distinct from the internal
COLI_OMP_TUNED re-exec sentinel);
- a one-line stderr breadcrumb is printed before the re-exec
("[OMP] hot-thread tuning: re-exec once (COLI_NO_OMP_TUNE=1 to skip)")
so the self-re-exec is self-documenting;
- execv only returns on failure, so a perror() now follows it
("execv self-reexec failed, running untuned") — a container without
/proc or a deleted inode falls back visibly instead of silently
losing the ~3.2x.
Fully opt-out / overridable and lossless:
- explicit OMP_WAIT_POLICY / GOMP_SPINCOUNT / OMP_PROC_BIND / OMP_DYNAMIC
in the environment win (overwrite=0);
- pre-setting COLI_OMP_TUNED=1 skips the re-exec entirely;
- COLI_NO_OMP_TUNE=1 disables the tuning path entirely;
- guarded by __linux__ (no-op re-exec elsewhere; the setenv defaults
still apply for any later-initialising runtime).
Claude-Session: https://claude.ai/code/session_01DS7oc65c5RdA9V99otRCwt
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The default ./glm mode validates against ref_glm.json, which is generated
from glm_tiny (a random tiny model). Pointed at the 744B GLM-5.2 it feeds
the model tiny-vocab prompt tokens and compares against tiny references —
0/20 and a degenerate loop, guaranteed on EVERY platform (x86/ARM/Metal),
which reads as an engine bug (it isn't). Detect the mismatch (real vocab +
tiny-range oracle ids) and print the correct commands instead. REF_FORCE=1
overrides. The engine self-test (SNAP=./glm_tiny TF=1) is unaffected.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New byte-level GBNF-subset engine (c/grammar.h: parser + set-of-stacks PDA
walker) wired into spec_decode as a third draft source ("metodo F"), tried
before MTP/n-gram. Wherever the grammar admits exactly one legal byte, the
forced span is tokenized and injected as drafts; the existing batch-union
verification confirms them, so a wrong or out-of-sync grammar can never
change the output. Lazy arming skips preambles; adaptive guard (same
pattern as MTP) disables the source below 50% acceptance; grammar-accepted
tokens no longer pollute the MTP acceptance counter.
GRAMMAR=file.gbnf enables it in run and serve modes (also with DRAFT=0 and
with the int4 MTP head from #8); GRAMMAR_DRAFT=n caps the span (default 24).
Measured on M3 Max / int8-MTP container, greedy, MTP=0 DRAFT=0, NDJSON
classification: 0.37 -> 0.50 tok/s (1.60 tok/forward, 81 fw per 130 tok),
100% acceptance (48/48), output byte-identical to baseline.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Model files come from untrusted mirrors. One choke point in load_cfg bounds
every dimension (hostile config.json now dies with a clear message instead
of reaching allocation sites); falloc guards the n*sizeof(float) wrap.
No hot-path calloc churn — the report's patch was declined in review, the
valid kernel of the idea is taken here.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
serve mode persists the compressed MLA KV-cache incrementally after every
turn (~182 KB/token appended, header count written last = crash-safe) and
resumes it at startup: the model remembers the whole conversation and zero
re-prefill happens. :reset and context-full restarts truncate the file.
The MTP layer's KV row is not saved; kv_start=-1 re-arms its decode window.
Validated: split-session answer byte-identical to an uninterrupted session
(tiny oracle, TEMP=0), and on the real 744B model a restarted chat resumed
58 tokens in 0.0s and recalled a fact from the previous session while
prefilling only the new question.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Measured on GLM-5.2 (48 tok, greedy): next-layer routing is 71.6% predictable
one full layer ahead (79.4% skipping attention only; 41.3% previous-token).
PILOT=1 issues next-layer expert WILLNEED from a dedicated I/O thread while
the current layer computes — inline fadvise BLOCKS ~0.5ms/call on a saturated
disk queue (+92s/48 tok, measured), hence the lock-free ring + worker.
Neutral-in-noise on this dev box (disk already ~80% duty); expected to pay on
balanced machines (#12: 43% disk / 46% matmul) — opt-in, default off.
cap_for_ram now RAISES the LRU cap up to the RAM budget (ceiling n_experts,
CAP_RAISE=0 opt-out): big-RAM machines were silently running with cap=8
(#12: 128GB box using 22GB of a 110GB budget; #13: 92GB box, same).
DRAFT=3 on cold cache measured locally: 1399s vs 880s baseline for the same
48 tokens (acceptance 16%, experts/token 1809 vs 800) — confirms #8; DRAFT
re-evaluation belongs to warm-cache serve sessions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- dot_i8i8/dot_i4i8: guarded __AVX512VNNI__ branches using vpdpbusd
(64 bytes/iter, s32 accumulation, no 16-bit intermediate). AVX-512 has
no vpsignb, so the sign trick becomes abs(w) + mask-negate on x.
- dot_i4i8: nibble unpack stays 256-bit; activation blocks are permuted
to match the per-lane value order instead of re-sorting weights.
- The int4 IDOT gate (was hardcoded S>=2, from AVX2 measurements) becomes
g_i4s with a VNNI-aware default of 1: single-row decode goes through the
integer path where it now pays. I4S=n overrides for A/B; I4S=2 restores
the old behavior exactly. AVX2/NEON/scalar builds are unchanged.
- Startup banner reports the active idot kernel (avx512-vnni/avx2/neon/scalar).
Measured on Xeon 6 (24C, AVX512-VNNI), DDR5-5600, ~19.8 GB/s NVMe RAID0,
gcc 13.3 -march=native: 256-tok greedy decode 1.72 -> 2.08 tok/s (+21%),
expert-matmul -24%; second prompt +20%. Integer-path outputs bit-identical
(associative s32 adds, same no-saturation bounds); S=1 int4 adopts the same
int8-activation tradeoff IDOT already makes for S>=2.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
- Faithful to the official modeling (transformers glm_moe_dsa): q from the
q_a latent via wq_b (32 heads x 128), k = LayerNorm(wk(h)) shared across
heads, interleaved RoPE on the first 64 dims, ReLU(q.k/sqrt(128)) weighted
by weights_proj(h)/sqrt(32), causal top-2048 per query.
- 'full' layers compute the selection (+ maintain the indexer k-cache from
token 0); 'shared' layers reuse it (IndexShare, index_topk_freq=4).
- Selection restricts both attention paths (absorbed decode + prefill
reconstruction). MTP row stays dense.
- Auto-detected like MTP: if out-idx-* weights are present for all full
layers, DSA arms itself; DSA=0 disables; DSA_FORCE/DSA_TOPK for testing.
- Validated on the tiny oracle (which ships indexer weights): selection
machinery forced on with keep=all keys reproduces dense attention exactly
(TF 32/32, gen 20/20); sparse smoke runs clean; kill switch verified.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- Learning cache: expert usage persists in <SNAP>/.coli_usage across sessions
(atomic save every turn); at startup the hottest experts are auto-pinned in
RAM with half the expert budget (AUTOPIN=0 disables). The engine gets faster
the more you use it.
- Sampling: temperature + nucleus (official 1.0/0.95 defaults in chat; TEMP=0
= greedy). MTP/n-gram speculation stays lossless via rejection sampling
(accept draft w.p. p(draft); on reject resample with draft banned).
- coli: --temp flag.
- Converter: --indexer mode extracts DSA lightning-indexer weights
(resumable; needed for future sparse attention beyond 2048 ctx).
- pin_load/stats include the MTP row; usage histogram covers layer 78.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rodolfo Hansen