This website requires JavaScript.
2824552d95
chore(snapshot): capture DSP reliability and backend work
2026-07-18 09:18:00 +09:00
5600a6ee52
feat(llm): typical-p / XTC logit filters as ops + sampling -inf/AccType precision fixes
2026-07-14 18:38:44 +09:00
8cd2649a08
fix(llm): preserve BOS for formatted prompts
2026-07-10 12:44:16 +09:00
b420f08c6a
feat(llm): ADR-0107 INT8 quantized KV decode V2 — row-inline scales, 100% token match
2026-07-10 10:25:25 +09:00
d54e4fbe1d
fix(dsp): complete #52 — unseal on new-borrower view invalidation instead of demote-crash
2026-07-07 06:40:18 +09:00
e479f26fd3
test(dsp): multi-GPU shard-boundary sweep + migration isView diag
2026-07-07 01:19:16 +09:00
6fb68ca8e4
fix(dsp): order the view-input migration dup against the peer copy
2026-07-07 00:55:12 +09:00
6b3a44f4a5
fix(dsp): multi-GPU output migration preserves strided (view) output layout
2026-07-07 00:37:12 +09:00
c9d6eaa40b
chore(dsp): strip multi-GPU ground-truth diagnostics now that roots are fixed
2026-07-06 19:39:33 +09:00
9aa7d9309d
feat(dsp): multi-GPU op-segment sharding with per-device captured-graph replay
2026-07-06 19:22:55 +09:00
93d12d8aca
fix(mem): enforce Environment.maxDeviceMemory as a real per-process device budget (#58 )
2026-07-05 17:25:34 +09:00
cd427d73ac
fix(dsp): OOM-failover 700 poison — async-ordered failover lifecycle (#57 , closes #46 WS-O3)
2026-07-05 13:17:12 +09:00
6c182b3c22
perf(dsp): WS-N4 steady-state trims + explicit cuBLAS TF32 opt-in
2026-07-05 07:23:01 +09:00
654a94a2ec
fix(serde): instance-accurate opNum for legacy ops in asFlatNode
2026-07-05 07:23:01 +09:00
c414f35310
fix(dsp): skip view-alias slots in post-replay fixup re-execution
2026-07-05 07:23:00 +09:00
0c19dd8999
fix(dsp): stale pointers in cached plans after donor teardown (#54 family)
2026-07-05 05:30:20 +09:00
44867e6f1c
docs(dsp): validation playbook — test tiers, change-type mapping, flaky/batch-only protocol (#29 )
2026-07-04 18:24:42 +09:00
b3d14d8923
fix(dsp): cast-cache skip-assign ABA across plan teardowns (#54 expression B)
2026-07-04 18:24:42 +09:00
400be2bd9f
fix(dsp): process-global deterministic cuBLAS window — off-thread GEMMs ran DEFAULT/TF32 (#55 )
2026-07-04 17:01:12 +09:00
ff5589b441
test(dsp): tighten alias-test JIT_TOL 1e-3 -> 2e-4 abs (post-PEDANTIC recalibration)
2026-07-04 15:56:35 +09:00
374a69cd8d
fix(dsp): audit-driven merged-replay actuality + JIT-mode PEDANTIC cuBLAS (#53 both roots)
2026-07-04 15:50:12 +09:00
0e7ae21663
fix(dsp,import,llm): whisper merged-decoder import chain + CUDA gate all-green campaign
2026-07-04 13:57:50 +09:00
0765462560
fix(dsp): dangling view slots over post-exec-closed buffers (#52 ) + close-path lifetime contracts
2026-07-04 12:16:23 +09:00
332cf220ef
fix(dsp): broad-sync external inputs when a cached plan's weights rebind (#12 batch divergence)
2026-07-04 09:18:18 +09:00
676272c0a4
fix(dsp): detach workspace-backed external inputs at plan ingestion (mid-decode UAF SIGSEGV)
2026-07-04 08:09:42 +09:00
3a167f2d10
fix(dsp): BooleanNot opName 'bool_not' -> canonical 'boolean_not' (traits never resolved)
2026-07-04 07:51:30 +09:00
4ea735bd6f
fix(blas): stop stomping the native BLAS-serialization default at JVM init (E3b)
2026-07-04 07:18:11 +09:00
893441b3bb
perf(cpu): paged/cascade attention decode parallelism + scratch reuse (E4)
2026-07-04 06:15:22 +09:00
8e37d3ae67
fix(win-cuda): disable OOM killer for CI — it kills nvcc mid-compile, wasting the 6h budget
2026-07-04 06:04:57 +09:00
ee32801c0e
fix(win-cuda): guard windows.h with NOGDI/NOMINMAX — ERROR macro clobbered enum (layer 5)
2026-07-04 06:00:35 +09:00
11c77db955
fix(bench): --dsp-timing never emitted — enable TIMING diag category at detailed level
2026-07-03 21:24:42 +09:00
572420c744
fix(win-cuda): include <windows.h> for Win32 CreateThread block in NativeDynamicShapePlan_cuda.cu
2026-07-03 21:19:46 +09:00
1a14621cee
fix(win-cuda-12.9): install cuSPARSE redist (12.5.10.65) — same C1083 fix as 12.6
2026-07-03 18:42:39 +09:00
c315e6e206
fix(win-cuda): install cuSPARSE redist — fixes C1083 cusparse_v2.h not found
2026-07-03 18:41:27 +09:00
997c770be6
fix(win-ci): normalize OPENBLAS_PATH to forward slashes (backslashes were stripped)
2026-07-03 17:36:11 +09:00
373c89d831
fix(win-ci): always set OPENBLAS_PATH for windows-cpu — fixes "No BLAS backend"
2026-07-03 16:33:40 +09:00
5947b7f917
fix(win-ci): SCCACHE_IDLE_TIMEOUT=0 — stop sccache server idle-death on windows-cuda
2026-07-03 14:31:33 +09:00
37d5df7aca
fix(win-cuda): drop MSVC-illegal extern explicit instantiation in vlm_ops.cpp
2026-07-03 12:07:38 +09:00
7c09f354eb
fix(ci): exclude platform-tests from cross-platform deploy (unresolvable test-only dep)
2026-07-03 11:40:45 +09:00
a9f1833e82
fix(omnihub): correct module-info requires — samediff.import.* -> samediff.modelimport.*
2026-07-03 11:20:13 +09:00
3e78938007
docs(omnihub): correct comment — release+add-exports conflict hits JDK 11 too, not just JDK 16+
2026-07-03 11:05:25 +09:00
08001b830f
fix(omnihub): drop redundant java.nio add-exports/add-opens that broke JDK 17 build
2026-07-03 10:52:17 +09:00
78e00edba9
fix(ui): restore DSP-diagnostics + stats-storage classes wrongly deleted as "hallucinated"
2026-07-03 10:40:46 +09:00
2bec0c2195
chore(llm): warn when speculative decoding is requested but not yet wired
2026-07-03 10:15:00 +09:00
367722a11d
fix(dsp): frozen-pin liveness registry for out-of-order teardown
2026-07-03 10:14:04 +09:00
81ab32f15a
fix(ci): repair windows-cuda, cross-platform, compat, and mac MLX/Metal builds
2026-07-03 10:12:00 +09:00
46e16112fe
fix(build): merge legacy -h helper into --helpers list (macOS mps-compile lost Metal)
2026-07-03 09:38:24 +09:00
c8dc3753a3
fix(ggml): make nd4j-ggml backend-agnostic (nd4j-api only, no test deps)
2026-07-03 08:40:52 +09:00
aad5c47674
fix(build): unbreak Windows CUDA link + centos-compat OpenBLAS detection
2026-07-03 08:31:56 +09:00
5ededbe232
fix(build): declare nd4j-common + slf4j-api deps in nd4j-presets-common
2026-07-03 05:36:00 +09:00
1e2c79df74
fix(ci): make accelerator build/smoke workflows dispatch-ready
2026-07-03 05:09:38 +09:00
8934a50c8d
fix(cpu): correct oneDNN/CPU ops on non-contiguous views (VLM vision-encoder corruption)
2026-07-03 04:43:52 +09:00
47de99e602
feat(accelerators): multi-backend enablement + islands/gaps mixed-runtime replay + CI tiers
2026-07-02 19:24:17 +09:00
3315472360
diag(dsp): route degraded-mode + wide-row-truncation through DspDiagnostics (F2, R2)
2026-07-02 17:56:48 +09:00
e42a948656
perf(dsp): per-token plan-cache key hashes only dynamic-shape ext inputs (D1a)
2026-07-02 16:05:37 +09:00
3edd8459b8
fix(dsp): make debug op-I/O value-dump view-safe (skip host preview for views)
2026-07-02 15:39:51 +09:00
18776e058b
fix(dsp): make KV in-place write capture-safe (no host cast/alloc mid-capture)
2026-07-02 15:13:50 +09:00
75f0808ce6
fix(triton): add boundary masks to fused MLP emitters (OOB on M<blockM)
2026-07-02 14:30:20 +09:00
e02e291179
perf(dsp): skip static dst/sizes H2D re-upload in launchBatchMemset (GPU)
2026-07-02 07:17:22 +09:00
2fc74000ce
perf(cpu): stop pinning MKL to 1 thread at library load
2026-07-02 06:55:37 +09:00
e49ac335b2
perf(dsp): gate per-token executor log.info behind isTraceEnabled
2026-07-02 06:48:06 +09:00
cdbb82ed81
perf: parallelize SDPA prefill across attention heads (CPU)
2026-07-01 21:14:32 +09:00
d52307d35e
snapshot: CPU BGE embedding perf (threading default, oneDNN/triton build, BlasHelper deadlock fix, SDPA per-row BLAS->SIMD, bench harness)
2026-07-01 21:01:30 +09:00
04088df4c0
fix: keep mixed external workspace gaps live
2026-06-30 18:56:22 +09:00
619f31b903
fix: reconcile live gap fast-path outputs
2026-06-30 17:26:57 +09:00
6ee580e443
fix: order merged capture validation replay
2026-06-30 17:06:49 +09:00
386acfcfe8
fix: refresh dynamic matmul cast caches for capture
2026-06-30 16:45:55 +09:00
5f78613768
fix: keep unsafe external-workspace gaps live for VLM decode
2026-06-30 16:18:07 +09:00
02aefda676
fix: stabilize Triton attention and ideal VLM benchmark
2026-06-30 10:49:17 +09:00
52b46758ee
fix: preserve VLM decode correctness and remove DSP workarounds
2026-06-30 07:06:07 +09:00
b4c7255d30
feat: working snapshot for graph ops and VLM decode fixes
2026-06-30 05:02:17 +09:00
158b30e698
feat: differentiable graph-analysis + GNN-codegen suite (branch checkpoint)
2026-06-29 09:22:58 +09:00
1735121865
fix: sync CUDA template-split build (CMake + buildnativeoperations) from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
cc464fe13c
fix: sync CudaDeviceConfig from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
027e25dd93
fix: sync array/memory infra (DataBuffer, ExtraArguments) from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
414f968e2b
fix: sync execution/legacy/loops infra from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
971241da90
fix: sync C++ helpers (Mmul/Blas/Cutlass/attention/ShapeBuffer) from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
6a1c0f6ca0
fix: sync C++ op implementations (attention/LLM/sampling) from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
9a8b477add
fix: sync cudnnUtils from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
bd34680686
fix: sync DSP/graph execution (C++) from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
18cd204dee
fix: sync Triton/NVRTC CUDA dispatch backend from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
c0be05eeea
fix: sync MmulHelper from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
4a1fb9949f
fix: sync DynamicShapePlanExecutor from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
1f0fdbda39
fix: sync DynamicShapePlanExecutor from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
e87fa9a79d
fix: sync benchmark helpers (Mmul/Pointers/Constant) from ag_new_release_updates_2
2026-06-19 08:52:22 +09:00
b3170f5f21
DSP dynamic-shape-plan execution + DataBuffer updates
2026-06-19 08:33:47 +09:00
ac096eb372
perf: core-aware MULTI_PASS chunk size (the real selective-rendering knob)
2026-06-19 06:15:25 +09:00
623f8dc29f
Snapshot: core-aware chunk target + extension-aware split_heavy_op
2026-06-18 21:40:36 +09:00
47dc38d704
perf: split heavy CUDA template-instantiation ops into per-type TUs
2026-06-18 20:58:25 +09:00
2cfdaeeeb5
DSP/graph + Triton CUDA dispatch refactor, platform splits, build auto-tune
2026-06-18 19:05:49 +09:00
74f46af4be
PR03: FlatBuffers schema update — add appended data offset/length to FlatArray
2026-06-18 06:35:34 +09:00
c6bd4b3953
fix: sync remaining DL4J convolution/gradient files from ag_new_release_updates_2
2026-06-18 06:08:20 +09:00
6a7303d208
fix: sync OnnxExportConfig from ag_new_release_updates_2
2026-06-18 06:08:05 +09:00
cf8bbc4b0c
feat: add OperationCategory, DeviceKey, GraphAnalysis from ag_new_release_updates_2
2026-06-18 06:08:01 +09:00
badcc5cb18
fix: sync Java op definitions from ag_new_release_updates_2
2026-06-18 06:07:48 +09:00
19b0da09d6
fix: sync ND4J API files from ag_new_release_updates_2
2026-06-18 06:07:33 +09:00
08f747156b
fix: sync graph files from ag_new_release_updates_2
2026-06-18 06:07:20 +09:00
1412018054
fix: sync remaining op files from ag_new_release_updates_2
2026-06-18 06:07:09 +09:00
3efd090c79
fix: sync helper headers from ag_new_release_updates_2
2026-06-18 06:06:55 +09:00
664ed790c0
fix: sync system config and Environment files from ag_new_release_updates_2
2026-06-18 06:06:45 +09:00
agibsonccc