Ettore Di Giacinto ccf75d1dcd docs(paged): record P1 bf16-stream landing (GO) ...

P1 of the EXECUTION_REARCH_SCOPE additive program landed: LLAMA_BF16_STREAM
(default-off) bf16-resident residual-segment executor for the q36 MoE model's
projection boundaries.

- EXECUTION_REARCH_SCOPE.md: dated "P1 RESULT" subsection (P0 kill-gate GO,
  full build-out deltas, KL, correctness gates, honest magnitude, provenance).
- PARITY_HANDOFF.md: chronology note (verdict, engagement, prefill/KL numbers,
  fork commits, deferred-not-failed measurements).

Key reframe recorded: q36 GDN/attention projections are BF16 weights (not
NVFP4), so bf16-stream is a MoE-model prefill lever; the dense model quantizes
those projections to NVFP4 and engages nothing (stays bit-identical). Prefill
MoE @512 +1.99% (reproducible, at noise floor), KL delta -0.00052 (KL-improving),
all md5 + test-backend-ops gates green. Fork HEAD 653bb2f3d, tree 6cf1523047.

Assisted-by: Claude:opus-4.8 [Claude Code]
Signed-off-by: Ettore Di Giacinto <mudler@localai.io>

2026-07-02 14:34:26 +00:00

..

ACCELERATOR_PORTING_SCOPE.md

docs(paged): scope porting the portable benefits to Metal/SYCL/Vulkan (+ROCm)

2026-06-28 08:34:32 +00:00

BENCHMARK.md

docs(paged): record phases 112-140 + series trim decision

2026-07-02 10:16:53 +00:00

DECODE_SERVING_SCOPE.md

docs(paged): record padded/fixed-slot decode shape as tested-and-rejected

2026-06-28 20:47:43 +00:00

EXECUTION_REARCH_SCOPE.md

docs(paged): record P1 bf16-stream landing (GO)

2026-07-02 14:34:26 +00:00

final_benchmark.csv

paged: drop bf16-tau (patch 0026), subsumed by decode fusions (tau=100000 flat, zero speed benefit)

2026-06-28 16:06:06 +00:00

GB10_PARITY_PHASE0_RESULTS.md

docs(paged): record BF16 F32 output broader serving phase

2026-07-01 13:26:50 +00:00

GB10_PARITY_REOPEN_SPEC.md

docs(paged): scope GB10 parity reopen plan

2026-06-30 15:44:11 +00:00

GDN_SHARED_AI_COST_MODEL.md

docs(paged): reject GDN global Ai32 prototype

2026-07-01 01:51:53 +00:00

LOCALAI_LLAMACPP_BACKEND_PLAN.md

chore(paged): keep patches/ patch-only; README to backend root, docs to docs/

2026-06-27 13:20:05 +00:00

PAGED_BITEXACT_NOTE.md

chore(paged): keep patches/ patch-only; README to backend root, docs to docs/

2026-06-27 13:20:05 +00:00

paged-burst-bench.cpp

chore(paged): keep patches/ patch-only; README to backend root, docs to docs/

2026-06-27 13:20:05 +00:00

paged-reclaim-unit.cpp

chore(paged): keep patches/ patch-only; README to backend root, docs to docs/

2026-06-27 13:20:05 +00:00

PARITY_HANDOFF.md

docs(paged): record P1 bf16-stream landing (GO)

2026-07-02 14:34:26 +00:00

PATCH_MAINTENANCE.md

docs(paged): record patch mirror readiness phase

2026-07-01 13:11:57 +00:00

PREFILL_GEMM_RESULTS.md

feat(paged): FP4 prefill large-M dequant->bf16 cuBLAS scaffold (patch 0033, default-off)

2026-06-28 17:42:15 +00:00

PREFILL_GEMM_SCOPE.md

docs(paged): scope the large-M NVFP4 prefill GEMM lever (design only)

2026-06-28 16:42:23 +00:00

qwen36_decode_overview.png

docs(paged): add the bf16-tau opt-in line to the decode plots

2026-06-27 22:25:02 +00:00

qwen36_dense_decode_vs_npl.png

docs(paged): add the bf16-tau opt-in line to the decode plots

2026-06-27 22:25:02 +00:00

qwen36_moe_decode_vs_npl.png

docs(paged): add the bf16-tau opt-in line to the decode plots

2026-06-27 22:25:02 +00:00

TENSORCORE_GDN_BUILD_PLAN.md

docs(paged): record GDN tensor-core revalidation phase

2026-07-01 14:05:20 +00:00

TENSORCORE_GDN_SCOPE.md

docs(paged): scope tensor-core (mma) chunked GDN prefill kernel

2026-06-28 17:23:51 +00:00

UPSTREAM_LAYER2_SCOPE.md

docs(paged): scope porting the portable benefits to Metal/SYCL/Vulkan (+ROCm)

2026-06-28 08:34:32 +00:00

VLLM_PARITY_FINAL.md

docs(paged): profile MTP graph reuse loss

2026-07-01 02:32:49 +00:00

VLLM_PARITY_LEVER_MAP.md

docs(paged): record datacenter Blackwell readiness phase

2026-07-01 14:28:41 +00:00