1.57.0 (2026-07-04)

Features

• cmd: bench_train -seed flag + exact loss bit patterns (T135.5) (d125201)
• docs: add EXTRA_ENV_PLACEHOLDER marker to validate-arm64 manifest (2fff208)
• scripts: plumb extra env vars through dgx-validate for env-gated fixtures (40b6653)
• timeseries: seedable weight init + GPU determinism double-run proof (T135.5) (0f6a316)

Bug Fixes

• attention: replay-stable FusedSDPA scratch (T133.2) (5093a9b)
• attention: stream-order flash-decode on the engine stream (T133.1) (db8c7b9)
• ci: add least-privilege permissions blocks to PR-triggered workflows (5ebc449)
• ci: pin mutable govulncheck/pip installs to exact versions (CICD-2) (3154617)
• cuda: bump ztensor to v1.19.1 and guard darwin dlopen probe (T137.1) (215fcef)
• cuda: guard FP16 graceful test from poisoning the CUDA context (T135.1) (047f2cc)
• cuda: guard sgemv_m1 float4 vectorized load on row pointer alignment (94834c4)
• cuda: include cooperative_groups/reduce.h in gemv_q4k_sm121 (d1dd5aa)
• cuda: port ztensor-required kernels missing from the fork (.so drift) (46c412c)
• cuda: port ztensor's flash_decode.cu into zerfoo's kernel fork (3c02533)
• cuda: remove CWD-relative dlopen candidates for the kernel library (CUDA-1, T141.1) (3f0d2fd)
• distributed: wire mTLS into worker gRPC server, refuse non-loopback without TLS (d6b2562)
• gguf: bounds-check tensor offset before mmap slice (F2, T139.2) (ab01d5d)
• gguf: check-before-multiply on tensor element-count overflow (F1, T139.1) (285fafe)
• gguf: restore missing math import in loader_mmap.go and split_file.go (ca1eb41)
• inference: serialize SpeculativeGenerate through the graph mutex (5e409a9)
• layers: FFN GELU backward always differentiated the SwiGLU branch (dd46acd)
• scripts: fail loudly on zero matched tests (T135.1) (9c8e4e4)
• serve: bound error-metric label cardinality via route normalization (9922a0b)
• serve: close model-delete TOCTOU and WaitGroup-misuse race (CONC-H2) (dcf1c90)
• serve: validate LoRA adapter name against traversal (SERVE-2) (3ad21c3)
• training: allocation-stable cached loss seed (non-arena device storage) (42b87a9)
• training: remove #878 containment gate (T133.4) (9968843)

1.56.0 (2026-07-03)

Features

• bench: add validate-arm64 Spark pod manifest (T131.1) (f95633a)
• scripts: -pkgs override for targeted GPU test scoping (T131.3) (2b96e9e)
• scripts: add -delete mode to dgx-validate for pod cleanup (T131.3) (3135db9)
• scripts: add dgx-validate.sh one-command GPU validation submit (T131.1) (9b500f6)
• scripts: committed in-pod validation stage dgx-validate-inpod.sh (T131.3) (518ef61)
• training: loud-fail gate on capture-replay training pending zerfoo#878 (T129.2) (72cfb48)

Bug Fixes

• bench: drop unprovisioned hostPath mounts from validate-arm64 pod (T131.3) (2b79ca4)
• bench: parameterize test scope via TEST_PKGS substitution (T131.3) (697ea89)
• bench: purego package set for validate-arm64 env (T131.3) (1e82694)
• bench: single-line bootstrap args -- Spark YAML drops block scalars (T131.3) (14c3141)
• cuda: skip dlsym invalid-handle test on linux -- handle 0 is RTLD_DEFAULT on glibc (T131.3) (7e29ef6)
• scripts: dgx-validate diagnosis hardening -- image pre-pull, events on failure, -keep flag (T131.3) (d6e8b0d)
• scripts: in-pod CGo header paths + QUALITY.md vet policy for purego warnings (T131.3) (7b6daa3)
• scripts: scope standing gate to purego GPU path; -tags cuda tracked in #921 (T131.3) (87465c5)

1.55.1 (2026-06-30)

Bug Fixes

• timeseries: seed iTransformer init for deterministic gradient checks (7115733)
• training: reserve captured-graph arena footprint under capture-replay (ztensor#167) (6cb6c02)

1.55.0 (2026-06-21)

Features

• regularization: add optional seeded RNG to Dropout and FeatureDropout (ec7dc54)
• regularization: capture-safe engine-op dropout path (WithEngineDropout) (8cedbc1)

1.54.0 (2026-06-18)

Features

• bench: bench_gemm + ltx2-fp8-spike manifest (E127/T127.1.0b) (92909be)
• layers: Conv3d + ConvTranspose3d forward (E127/T127.4.2-4.3) (374b807)
• layers: GroupNormalization layer + LTX-2 VAE decoder skeleton (E127) (7a5ffc8)

Bug Fixes

• optimizer: keep AdamW grad-norm reduction on-device (T11.8) (39750c7)
• parity: build GPU parity container with CGO, no cuda tag (#902) (58731e9)

1.53.1 (2026-06-17)

Bug Fixes

• training: bf16 device-resident gradient accumulation uses the engine (f8570db)

1.53.0 (2026-06-17)

Features

• layers/core: use MatMulTransposeA for Linear dW gradient (bf16 GPU path) (a56365d)

1.52.0 (2026-06-16)

Features

• layers: make Softmax, Linear, Conv1D, VarSelection, CE-loss bf16-instantiable (138c233)

Bug Fixes

• optimizer: keep AdamW hyperparameters in float64 for reduced-precision T (9083452)

1.51.0 (2026-06-16)

Features

• optimizer: on-device AdamW step when engine + params are GPU-resident (ADR 070, ADR 075 L1) (ebc0186)

1.50.3 (2026-06-16)

Performance

• linear: device-resident dInput via MatMulTransposeB (ADR 075 L1) (cf830bd)