Commit Graph
873 Commits
Author SHA1 Message Date
Ettore Di Giacinto 67d2c4c9d4 docs(paged): record BF16 GDN state cache phase
Record Phase81 default-off BF16 persistent S-cache results, including md5 drift, op gates, decode profile, and KL smoke. Scope Phase82 as full f16-reference KL plus serving A/B before patch-series promotion.

Assisted-by: Codex:gpt-5
2026-07-01 16:26:09 +00:00
Ettore Di Giacinto d091eb30f2 docs(paged): record GDN identity shortcut phase
Assisted-by: Codex:gpt-5
2026-07-01 15:46:41 +00:00
Ettore Di Giacinto bbfaa66f02 docs(paged): record GDN BV32 decode A/B phase
Assisted-by: Codex:gpt-5
2026-07-01 15:35:06 +00:00
Ettore Di Giacinto 04ed7fe52f docs(paged): record GDN launch sweep phase
Assisted-by: Codex:gpt-5
2026-07-01 15:13:45 +00:00
Ettore Di Giacinto a9454b45c8 docs(paged): record MoE decode-only profile phase
Assisted-by: Codex:gpt-5
2026-07-01 15:05:45 +00:00
Ettore Di Giacinto f21b393746 docs(paged): record current MoE graph profile phase
Assisted-by: Codex:gpt-5
2026-07-01 14:56:39 +00:00
Ettore Di Giacinto 26a41fad1a docs(paged): record post-PoC GDN audit phase
Assisted-by: Codex:gpt-5
2026-07-01 14:44:17 +00:00
Ettore Di Giacinto 5369219729 docs(paged): record GDN blocked-solve PoC phase
Assisted-by: Codex:gpt-5
2026-07-01 14:39:09 +00:00
Ettore Di Giacinto eb82ff138f docs(paged): record datacenter Blackwell readiness phase
Assisted-by: Codex:gpt-5
2026-07-01 14:28:41 +00:00
Ettore Di Giacinto 2efb0ec362 docs(paged): record TTFT min32 serving phase
Assisted-by: Codex:gpt-5
2026-07-01 14:18:54 +00:00
Ettore Di Giacinto e5c5746c0a docs(paged): record GDN tensor-core revalidation phase
Assisted-by: Codex:gpt-5
2026-07-01 14:05:20 +00:00
Ettore Di Giacinto 6cf8b782d1 docs(paged): record BF16 F32 output broader serving phase
Assisted-by: Codex:gpt-5
2026-07-01 13:26:50 +00:00
Ettore Di Giacinto e573194799 docs(paged): record patch mirror readiness phase
Assisted-by: Codex:gpt-5
2026-07-01 13:11:57 +00:00
Ettore Di Giacinto 2b2b1f0b25 docs(paged): record BF16 F32 output dense serving phase
Assisted-by: Codex:gpt-5
2026-07-01 13:06:49 +00:00
Ettore Di Giacinto e67b329eb1 docs(paged): record BF16 cuBLAS F32 output phase
Assisted-by: Codex:gpt-5
2026-07-01 12:54:24 +00:00
Ettore Di Giacinto 60954d484a docs(paged): record quant kernel timing phase
Assisted-by: Codex:gpt-5
2026-07-01 12:45:19 +00:00
Ettore Di Giacinto 3fbdfc21c9 docs(paged): record quant trace phase
Assisted-by: Codex:gpt-5
2026-07-01 12:42:13 +00:00
Ettore Di Giacinto 55df9100dc docs(paged): record layout trace phase
Assisted-by: Codex:gpt-5
2026-07-01 12:32:05 +00:00
Ettore Di Giacinto 2e19e5c90f docs(paged): record prefill bucket attribution phase
Assisted-by: Codex:gpt-5
2026-07-01 12:20:42 +00:00
Ettore Di Giacinto 6a2618b6dc docs(paged): record MTP verify-cost rejection
Assisted-by: Codex:gpt-5
2026-07-01 11:51:29 +00:00
Ettore Di Giacinto f7d76389b0 docs(paged): record W4A16 direct activation rejection
Assisted-by: Codex:gpt-5
2026-07-01 11:28:11 +00:00
Ettore Di Giacinto ef578866c8 docs(paged): scope W4A16 direct activation experiment
Assisted-by: Codex:gpt-5
2026-07-01 10:59:56 +00:00
Ettore Di Giacinto fc5d5e4ff3 docs(paged): profile current W4A16 prefill
Assisted-by: Codex:gpt-5
2026-07-01 10:56:48 +00:00
Ettore Di Giacinto ef7dbfa5f7 docs(paged): compare MoE min32 against vLLM
Assisted-by: Codex:gpt-5
2026-07-01 10:46:32 +00:00
Ettore Di Giacinto c41d1a5b4f docs(paged): record waiting-threshold TTFT defer
Record Phase58 prompt-backlog threshold A/B, DGX gates, MoE and dense serving results, and the repeat-before-default decision.

Assisted-by: Codex:gpt-5
2026-07-01 10:31:09 +00:00
Ettore Di Giacinto 9be291e6b0 docs(paged): reject capped TTFT defer sweep
Record Phase57 capped TTFT prefill-first sweep, DGX gates, and the decision to keep the cap as an A/B knob rather than a parity path.

Assisted-by: Codex:gpt-5
2026-07-01 10:18:41 +00:00
Ettore Di Giacinto 902bcc7717 docs(paged): validate TTFT prefill-first A/B
Record Phase56 MoE and lower-concurrency validation for the TTFT prefill-first policy, including DGX gates and the opt-in-only decision.

Assisted-by: Codex:gpt-5
2026-07-01 10:05:23 +00:00
Ettore Di Giacinto 999cf09532 docs(paged): record TTFT prefill-first A/B
Record Phase55 default-off scheduler A/B, DGX md5/op gates, dense serving results, and the pending fork push/mirror status.

Assisted-by: Codex:gpt-5
2026-07-01 09:57:55 +00:00
Ettore Di Giacinto 3dbf34e739 docs(paged): record admission histogram trace
Record Phase54 trace-only histogram work, DGX md5/op gates, dense serving histogram evidence, and the next scheduler decision.

Assisted-by: Codex:gpt-5
2026-07-01 09:40:50 +00:00
Ettore Di Giacinto 347a5c05bd docs(paged): reject admission budget sweep
Assisted-by: Codex:gpt-5
2026-07-01 09:27:20 +00:00
Ettore Di Giacinto 2aa76702df docs(paged): record dense admission trace
Assisted-by: Codex:gpt-5
2026-07-01 09:18:43 +00:00
Ettore Di Giacinto b5f65152e2 docs(paged): record serving admission trace
Assisted-by: Codex:gpt-5
2026-07-01 09:08:42 +00:00
Ettore Di Giacinto c299dcd231 docs(paged): record dense true decode profile
Assisted-by: Codex:gpt-5
2026-07-01 08:55:23 +00:00
Ettore Di Giacinto cd59e5d61f fix(paged): scrub harness vars for vllm serve
Assisted-by: Codex:gpt-5
2026-07-01 08:23:05 +00:00
Ettore Di Giacinto 96825a224e docs(paged): record dense serving snapshot
Assisted-by: Codex:gpt-5
2026-07-01 08:20:26 +00:00
Ettore Di Giacinto 440129c98e fix(paged): harden serving snapshot readiness
Assisted-by: Codex:gpt-5
2026-07-01 08:07:48 +00:00
Ettore Di Giacinto e69ee0e867 feat(paged): parameterize served model name
Assisted-by: Codex:gpt-5
2026-07-01 07:50:19 +00:00
Ettore Di Giacinto 2a0fc0f4b9 docs(paged): record inference gate guard
Assisted-by: Codex:gpt-5
2026-07-01 07:45:52 +00:00
Ettore Di Giacinto ae8284f5fb feat(paged): parameterize vllm serving snapshot
Assisted-by: Codex:gpt-5
2026-07-01 07:41:55 +00:00
Ettore Di Giacinto ecaf406c0b docs(paged): reject persistent gate fusion shortcut
Assisted-by: Codex:gpt-5
2026-07-01 07:34:27 +00:00
Ettore Di Giacinto b9eff5bca3 docs(paged): reconcile next parity target
Assisted-by: Codex:gpt-5
2026-07-01 07:31:26 +00:00
Ettore Di Giacinto aa848d5afb docs(paged): record low-concurrency serving check
Assisted-by: Codex:gpt-5
2026-07-01 07:24:28 +00:00
Ettore Di Giacinto d44e164c96 docs(paged): record max-concurrency parity check
Assisted-by: Codex:gpt-5
2026-07-01 07:13:48 +00:00
Ettore Di Giacinto 52c11b1ce5 docs(paged): reject graph-time gate fusion shortcut
Assisted-by: Codex:gpt-5
2026-07-01 06:56:01 +00:00
Ettore Di Giacinto 5354adcffb docs(paged): scope gate projection policy
Assisted-by: Codex:gpt-5
2026-07-01 06:50:19 +00:00
Ettore Di Giacinto 9f75da01f9 feat(paged): add cublas tensor-name trace patch
Add patch 0063 extending LLAMA_CUBLAS_ROUTE_TRACE with src0/src1/dst tensor names.

Record Phase 37 gates and the conclusion that SGEMM traces to MoE gate tensors.

Assisted-by: Codex:gpt-5
2026-07-01 06:41:00 +00:00
Ettore Di Giacinto fbdc200886 feat(paged): add cublas route trace patch
Add patch 0062 with default-off LLAMA_CUBLAS_ROUTE_TRACE instrumentation for generic cuBLAS MUL_MAT subroutes.

Record Phase 36 DGX gates, serving trace results, and the next projection follow-up scope.

Assisted-by: Codex:gpt-5
2026-07-01 06:24:46 +00:00
Ettore Di Giacinto 49cce0b5a2 feat(paged): add mul mat route trace patch
Add LocalAI patch 0061 from the llama.cpp fork and record Phase 35 gates, serving route counts, and the updated patch mirror invariant.

Assisted-by: Codex:gpt-5
2026-07-01 05:52:09 +00:00
Ettore Di Giacinto ba1979a689 feat(paged): add moe mmid route trace patch
Add LocalAI patch 0060 from the llama.cpp fork and record Phase 34 gates, serving route counts, and the updated patch mirror invariant.

Assisted-by: Codex:gpt-5
2026-07-01 05:37:53 +00:00
Ettore Di Giacinto 7665422bfa feat(paged): add moe small-m mmq tile policy gate
Assisted-by: Codex:gpt-5
2026-07-01 05:20:18 +00:00