Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

Update Feather H200 training runtime image

fix: use wte.weight.shape instead of undefined config var

fix: lm_head init std=0.02 (was 0.001 — pathological at V=65k) + retina HF cache

fix: lm_head init std=0.02 (was 0.001 — pathological at V=65k) + retina HF cache

fix(mid_val): use prepare.get_token_bytes(device=cuda) + model(x,y,reduction=none) API

feat(sdr_retina): streaming Nemotron path when HYDRA_USE_NEMOTRON=1

fix(tokenizer): use rustbpe.train_from_iterator API; bump vocab to 65536

fix(tokenizer): use rustbpe.train_from_iterator API; bump vocab to 65536

perf(nemotron): 2-stage prefetch pipeline (HF→tokenizer→packer) zero-tps-loss

feat(nemotron): streaming pretraining loader for Specialized-v1.1 (Super3 recipe)

feat(nemotron): streaming pretraining loader for Specialized-v1.1 (Super3 recipe)

feat: add ppl to per-step log + MID_VAL every 500 steps for learnability visibility

perf(htm): T9 input-tile memcpy_async to cluster smem (bandwidth reduction)

fix(htm): add cluster.sync between Stage A writes and next-timestep reads (T8 bimodal fix)

feat(triton-cache): wire setup/teardown into entrypoint

feat(triton-cache): HF Hub-backed compilation cache persistence

perf(htm): cluster distributed shared memory for inhib_thr/boost/active_duty (T8)

fix(htm): slice regions[:B] to handle eval batch size differing from training

fix(htm): set NON_PORTABLE_CLUSTER_SIZE_ALLOWED=1 for cluster_size=16

perf(htm): Hopper cluster::sync hardware barrier + sm_90a + cluster launch attr

perf(htm): Hopper cluster::sync hardware barrier + sm_90a + cluster launch attr

perf(htm): Hopper cluster::sync hardware barrier + sm_90a + cluster launch attr

perf(htm): Hopper cluster::sync hardware barrier + sm_90a + cluster launch attr

perf(htm): DLB software grid barrier + non-cooperative launch (lifts 132-SM cap)

perf(htm): DLB software grid barrier + non-cooperative launch (lifts 132-SM cap)

fix(htm): drop py.allow_threads (raw ptrs not Send)

perf(htm): batched cooperative kernel — B=8 regions in ONE launch via blockIdx.y indexing

perf(htm): batched cooperative kernel — B=8 regions in ONE launch via blockIdx.y indexing

perf(htm): batched cooperative kernel — B=8 regions in ONE launch via blockIdx.y indexing

perf(htm): batched cooperative kernel — B=8 regions in ONE launch via blockIdx.y indexing

perf(htm): thread-pool dispatch of B regions concurrently + HTM_FUSED_GRID_CAP=16 for kernel concurrency

perf(htm): remove per-call dev.sync, cache SDR, device_sync once per step

perf(htm): remove per-call dev.sync, cache SDR, device_sync once per step

perf(htm): remove per-call dev.sync, cache SDR, device_sync once per step

fix: dD_acc sum via axis=0 not scalar.reshape(1)

patch: 7 triton 3.4 compat fixes (3 tl.sum + 3 bwd desc.dtype + 3 fwd desc.dtype)

patch: add bf16 casts on TMA descriptor stores (triton 3.4 strict dtype match)

compat: triton 3.4 M,N,K>=16 patches for mamba3 SISO kernels (tl.dot -> tl.sum for sum-reductions)