- register online_ppo train type backed by PPOStrategy: token-level clipped surrogate over GAE advantages plus masked value regression against rollout-pinned returns, with explained-variance metrics
- fold the reference-KL penalty (k3 estimator) into per-token rewards before GAE and pin advantages/returns on RolloutResult so replayed gradient steps optimize fixed targets
- add self-contained ValueModel critic with a zero-initialized value head and backbone warm-started from policy weights; AutoRegressiveLM stays untouched and trunk parity is pinned by tests
- step the critic's own optimizer outside the policy-version lock with the same max_grad_norm clipping as the policy
- persist critic state as value_model.pt/value_optimizer.pt checkpoint extras; resume restores it, fails loudly when missing, and the train.sh completeness check requires the extras for online_ppo configs
- extract shared rollout sequence/logprob helpers from GRPO (behavior unchanged) and add ppo_gamma/ppo_gae_lambda/ppo_vf_coef CLI options
- sliced M=2 and full M=5 lm_head projections may pick different GEMM kernels whose accumulation order differs in the last float32 bits, so torch.equal flakes by machine and thread count
- delete csrc/kernels/gemm.cu and swiglu.cu and drop their CMake and setup.py registration
- remove the ops wrappers plus backend/linear.py and backend/swiglu.py so Linear and MLP call F.linear directly
- drop the four gemm and swiglu kernel test files and prune the stale cuda_kernels.md sections
- add csrc/bench benchmarks for the remaining kernels: attention decode prefill paged decode paged prefill versus single-launch SDPA references, rotary versus the torch fallback, fp8 quantize and mm_fp8 versus torch baselines
- attention, rotary_emb, and fp8_ops kernels are unchanged
- add logits_positions to AutoRegressiveLM.forward, gathering rows before the final norm so the lm_head GEMM covers only the positions prefill samples from
- execute_prefill builds last_token_indices up front and passes them in, dropping the post-forward gather of a [tokens, vocab] tensor
- prefill graph warmup passes a single index; decode stays untouched (every row is sampled) and prefill itself runs eager, so graph capture is unaffected
- update the ragged-prefill fake to slice by the received index and add a packed-row exact-equality test
Benchmark: NVIDIA L20 (idle), CUDA 12.8, torch 2.11.0+cu128, 1.2B bf16 checkpoint, 512-token prompts, greedy; prefill B=32: 368.1 -> 323.5 ms (44.5k -> 50.6k tok/s, +13.8%), B=8: 89.3 -> 78.9 ms (+13.2%), B=1: 12.3 -> 11.4 ms (+7.9%); decode step unchanged; full suite: 897 passed
- add BatchedStreamCallback sink type: TaskManager resolves a decode step's (task_id, token) events under one lock and delivers each sink a single list instead of one call per token
- keep the plain Callable[[str]] callback contract: per-token callbacks still receive one call per event, and invoke_callback/cancel_task wrap single events for batched sinks
- collect aborted, text, and finish STOP events in the scheduler decode loop and dispatch once per step instead of once per token
- register one _ResultSink per generate call (replacing per-task closures) so GenerateResult takes its lock and wakes waiters once per step, with late-bind replay for tasks that start decoding before add_task returns their id
- apply GenerateResult batches under a single condition hold via append_batch; append delegates to it
- update engine test fakes to the batched contract and add coverage for event grouping, single-event dispatch, cancel STOP, and late-bind replay
Benchmark: NVIDIA L20 (idle), CUDA 12.8, torch 2.11.0+cu128, 1.2B bf16 checkpoint, prompt 512, 256 greedy tokens, CUDA graph on, serving-level decode, 3 trials
- batch 32: 7.808 -> 7.506 ms/token (4098 -> 4263 batch tok/s, +4.0%)
- batch 1/8: unchanged within noise (3.768 -> 3.797 / 4.699 -> 4.607 ms/token)
- full suite: 896 passed
- split-K removed entirely: tiled kernel walks K in one pass, no partials/semas workspace, no memset, single launch per call
- skinny GEMM (M<=8) dispatch table replaces the hand-written switch
- shape-driven four-family table replaces plan_gemm: wide-N (n>=4096) default {16,64,64,3,128} with BM=32 at M>16; narrow-N deep-K rings {16,32,256,2,64} while the grid fits one wave, {16,32,128,2,64} past it
- narrow-N is K-serial: widening the grid measurably does nothing (BN 64->32 ties, doubled m_tiles tie, kv at 4 blocks ties q/o at 24); deeper K chunks win until 72KB smem forces one CTA per SM and past one wave the 2-wave quantization loses to BK=128
- launch-check macros in common/launch.cuh; smem opt-in for the 72KB/60KB rings
- rename kernels/bf16_*.cu to gemm.cu/swiglu.cu; module names unchanged
- Python gate: lm_head (N>32768) falls back to cuBLAS, band narrows to M<=32
- drop the stale per-op benchmark narratives; fold the live numbers into cuda_kernels.md
Benchmark: NVIDIA L20 (sm_89, 92 SMs), CUDA 12.8, bf16, L2-thrash weight rotation, per-call medians at M=16: q/o 9.5us, kv 8.6us, gate/up 33.3us, down 33.7us (down -29% vs prior default). End-to-end 1B decode (gen 128, 3 trials, tokens/s vs cuBLAS): B=1 260 vs 252, B=8 1660 vs 1446, B=16 2464 vs 2437, B=32 3620 vs 3690. Prior split-K dispatch measured B=16 2243 / B=32 3393.
- make decode steady state self-validating: gate the token fill and sampling reuse on the decode cache's own task signature, drop TaskCacheManager.last_task_signature_matches whose req-index signature recycles with slot reuse
- extract PolicyVersionGuard (version protocol plus generation/weight mutex) and Stepper (shared one-token advancement) out of the scheduler, keeping its public API unchanged
- pin the input staging buffer only on CUDA devices so CPU-only workspaces allocate
- split KVCache into phase-specific PrefillKVCache/DecodeKVCache types selected by start_pos
- unify steady-state detection in TaskCacheManager
- guard decode steady-state reuse with the cached task signature so recycled req slots cannot replay a prior generation's tokens and positions
- collapse attention backend fwd_decode/fwd_prefill into a single subclass-owned forward with a shared _check_fwd guard
- fix thread-safety gap in weight update and validate prefill inputs before KV allocation
- centralize magic constants in InferenceConfig and align docs with behavior
- add build_engine() to astrai.inference.engine as the single load-place-wire path for InferenceEngine, accepting a checkpoint path or live model/tokenizer plus passthrough engine kwargs
- migrate the server lifespan, generate CLI, humaneval/ifeval evals, and all three demos to build_engine; app._create_engine collapses into a direct call
- export build_engine from astrai and astrai.inference
- parameterize the autoregressive demo with --prompt one-shot continuation plus model path and sampling knobs, exiting cleanly on !exit or EOF
- cover the composition root with unit tests for live-object assembly, kwargs passthrough, and argument validation
- extract shared helpers for dataset writers, scheduler construction, thread interleaving, hf roundtrips, and moe configs
- remove about 20 cases whose only assertions were format checks, restated declarations, fake-taxonomy duplicates, or test-local scaffolding
- strengthen weak cases into exact reference comparisons, positional mask checks, and deterministic outcomes
- replace two schedule factory smoke tests with cosine/sgdr formula assertions
- delete root-level CLI tests whose merge-priority facts are covered by tests/config/test_cli.py
- suite shrinks from 857 to 826 items; ruff format, import order, and pytest all green
- Add astrai/config/cli.py: OptSpec tables plus apply_specs infer click types and defaults from config fields, covering Optional[X], Union[X, None], PEP 604 X | None, stringified PEP 563 annotations, bool flag pairs, and repeatable list options
- Move GroupedCommand/GroupedOption and the three-layer YAML merge (option defaults < YAML < explicit CLI) into the config package, adding unknown-key warning and mapping validation
- Replace ~420 lines of hand-written @opt decorators in scripts/tools/train.py with a 66-entry spec table; option names, defaults, flag styles, and YAML semantics verified unchanged
- Migrate scripts/tools/server.py to the same mechanism with its section binding, integer coercion, and dtype validation preserved locally
- Add tests/config/test_cli.py covering type inference across annotation styles, default overrides, flag pairs, merge precedence, scientific notation, and help ordering
- save an emergency checkpoint even when the signal is handled before the first optimizer step, where optimizer_step == last_ckpt_step used to skip the save entirely (CI race on slow cold-start runners)
- track a saved-this-session flag so interrupted runs always have at least one checkpoint, while normal zero-step completion keeps skipping the save
- shard the Muon Newton-Schulz orthogonalization over the FSDP mesh instead of partial local slices
- import HF checkpoints faithfully: per-head RoPE permutation for q/k projections and qk-norm, qwen3, shared experts, and qk-norm before RoPE (changes numerics for existing use_qk_norm checkpoints)
- make preprocessing and resume self-contained: backfill realigned bucket keys by semantics (masks ones, rest zeros) and snapshot tokenizer files into every checkpoint
- keep RL consistent: sync the offline GRPO old_model each optimizer step and validate online strategies through a public one-off-rollout hook that leaves the replay cache untouched
- fix streaming serving: withhold partial tool-call prefixes with a stream-end flush, stream tool-call arguments from the raw source span, and terminate SSE frames with a blank line
- fix sampling semantics: capture logprobs before top-k/top-p mutate logits in place and detect greedy pipelines polymorphically instead of isinstance bookkeeping
- re-register the linear family with the operator dispatcher (ASTR_OPS / op_backend / resolve)
- fix bf16 gemv misaligned-address faults and element mispairing for offset weights
- reject misaligned bf16_swiglu inputs with a clear error and fall back in the backend gate
- make the rollout reuse decision, validation, and return atomic under one policy snapshot
- add the documented post-scoring rollout version check
- derive live+1 under the scheduler lock in optimizer_step via apply_weight_update(None, ...)
- reject rollout_max_policy_lag below rollout_interval - 1 at config time
- sync gemv stream-test inputs before switching streams; drop dead loader imports
- replace the per-shape auto tables in the linear backend with an M-banded rule (M in [2,4] on compute capability 8.0+) that measured at the HBM bandwidth floor across every family, and fold the capability check into the capable guard
- drop the unreachable swiglu auto shape-table machinery so both backends share one env-mode ladder via the new dispatch.env_mode helper
- add __all__ across extension modules, name the rotary registration records, and unify typing to the typing-module style
- rewrite test_linear_dispatch.py around behavioral routing assertions and document the M-banded policy in the developer docs
- Benchmark: L20 SM89, Python dispatch overhead 2.9us to 1.5us, auto now covers every projection shape at M in [2,4].
- delete the warp-tiled kernel and both per-shape (N,K) selector tables; block size is 256 threads everywhere except M=8 with N*K <= 12 MiB, which keeps a 128-thread CTA
- HBM-streaming measurements (weight copies rotated through L2, the real decode regime) show the variants within ~3% on L20 because the kernel is bandwidth-bound; the retired tables were tuned against an L2-resident loop and sometimes picked the slowest variant ((2048,8192) M=8: coop128 6% slower than coop256)
- a shape no longer switches kernels (and accumulation order) with M, removing one shape-dependent nondeterminism source
- remove the stale split-K launcher comment
- move bf16_gemv.cu and bf16_swiglu.cu from csrc/kernels/gemv/ to csrc/kernels/ beside rotary_emb.cu; the family keeps no shared headers
- rename test_bf16_gemv_matches_half_cta_edge_bands to test_bf16_gemv_matches_m8_edge_bands and update docs/developer/cuda_kernels.md
Benchmark: L20 (sm_89), PyTorch 2.11.0+cu128, interleaved CUDA-event timing with rotated weight copies exceeding the 96MB L2; variant spread <=3% across 14 shapes x M in {1,2,4,8}, and the retained rule wins 5-9% at M=8 small weights ((512,3584), (1536,1536), (6912,1536))
- deepen common-shape BF16 GEMV tuning with warp-row tiling for LLaMA/Qwen2/GPT-NeoX/OPT decode projections
- add fused BF16 up/gate SwiGLU CUDA primitive with ASTRAI_SWIGLU=0/1/auto dispatch
- keep the unfused linear backend as the default path; auto enables no shape until per-architecture checkpoint gates pass
- fall back to the linear/torch chain when kernels are absent, on CPU, in training, or outside supported M/K/dtype shapes
- add gemv/swiglu benchmark scripts, dispatch and parity tests, and kernel documentation
Benchmark: NVIDIA L20 (sm_89), CUDA 12.8, PyTorch 2.11.0+cu128, idle GPU. AstrAI 1B config (24 layers, hidden 1536, vocab 100000), BF16, prompt 128, 32 greedy decode tokens, CUDA graphs enabled, A/B in separate interleaved processes (3 rounds, 8 trials each, medians). Default vs ASTRAI_SWIGLU=1 per generate call: batch 1 134.8->129.1 ms (+4.44%), batch 2 136.2->130.9 ms (+4.06%), batch 4 145.5->140.3 ms (+3.66%). Greedy output identical at batch 1, differs at batch 2/4, so auto stays unfused by default; kernelless fallback verified bit-identical greedy.
- reject prompts that encode to zero tokens in add_task instead of admitting a task whose prefill can never run, and surface empty-id run_batch calls as prompt_empty errors
- deliver the STOP stream callback when cancelling a live task so clients observe termination instead of hanging until socket timeout
- strip the torch.compile _orig_mod. prefix at every unwrap_model site and when loading checkpoints so FSDP state dicts and saved weights no longer leak the wrapper name into downstream keys
- reject online_* train strategies with nprocs > 1 at config validation time, explaining the NCCL all-gather deadlock they would otherwise hit mid-run
- apply the frequency penalty before temperature scaling (OpenAI semantics) so the penalty survives temperature=0 instead of being annihilated by the 1e8 logit blowup, and exclude penalty pipelines from the greedy fast path
- return logprobs from the raw pre-strategy distribution so they match training-side policy logprobs for PPO/GRPO importance ratios
Feed sampler-aligned behavior log-probabilities directly into online GRPO instead of allocating, synchronizing, and forwarding a duplicate old-policy model. Keep the old-model path as an offline compatibility fallback and validate supplied rollout tensors before loss computation.
Track a monotonic policy version across optimizer steps, scheduler updates, and rollout results. Serialize synchronous generation with weight acknowledgements and invalidate reusable prefix KV entries so cached samples remain attributable to the behavior policy that generated them.
- Write checkpoint payloads to a hidden sibling staging directory, add a versioned checksum manifest, fsync the completed payload, and publish it with an atomic rename
- Republishing an existing step retires the old payload under a hidden sibling name before the atomic rename, so re-runs into the same output directory replace the previous checkpoint instead of raising FileExistsError
- Keep legacy checkpoints loadable, add optional checksum verification, and align metric flushing with checkpoint publication
Co-authored-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
- Pack prompts with a shared prefix start and attention backend into one forward.
- Select per-request final logits from cumulative query lengths.
- Cover ragged tokens, logprobs, scheduling, and documentation.
- Drop the K % 2 entry rejection and the per-K if/else load-width branch: the weight stream now anchors uint4 loads at each row's first 16-byte-aligned address, with scalar head/tail sweeps covering at most 14 remainder elements, so any positive K and any storage offset is correct
- Keep one pure-uint4 loop (no branching inside the loop) for the production case where every x row base is 16-byte aligned (K % 8 == 0 with allocator-aligned tensors) and a scalar-x pairing loop only for unaligned K, where per-row uint4 loads are not addressable; measured cost of scalar x everywhere was up to 2.5x on multi-row shapes (down M=4 28.4us vs 11.3us)
- Remove the now-obsolete k_aligned axis and K divisibility gate from the linear dispatch spec since the primitive no longer rejects any K
- Add test coverage for unaligned K (7, 12, 100, 1534) at M=1 and M=3
Benchmark: 8x L20 (sm_89, CUDA 12.8), L2-resident microbench, 300 iters; hot path unchanged within noise vs the pure-uint4 kernel (q M=2 5.8us, down M=4 11.3us, lm M=1 391us); full gate green
- Replace per-element loads with 128-bit uint4 vectorized loads (8 halves per access), improving every measured shape: q/k/v at M=2 from 6.0us to 5.4us, q_proj speedup 2.28-2.45x, mlp_down at M=4 2.76x, lm_head at M=1 +6-8%
- Extend kernel M support from {1,2,4,8} to all M in 1-8 via new BLOCK_M cases 3,5,6,7, since cuBLAS wmma templates pad small M to 8/16 rows and waste compute
- Keep the auto-dispatch allowlist unchanged: a 64-step greedy-walk probe on the real decode path showed mlp_down (K=6912) divergence at step 1 and argmax flips for every candidate odd-M band, the same noise class already present in the merged M=2/4 entries, so no entry has the stability evidence the gate requires
- Rejected alternatives with measurements: split-K accumulation (k/v shapes regress 6.0us to 9.2us, code removed) and MMA tiles (small M is DRAM-bound at ~1 FLOP/byte vs the ~138 needed)
- Update test_gemv M-rejection case to M=9 and test_linear_dispatch multirow fallback to M=9 for the widened range
Benchmark: 8x L20 (sm_89, CUDA 12.8), single-GPU microbench, 200 iters after 20 warmup, weights L2-resident; q(1536x1536) M=3 8.9->5.3us, kv(256x1536) M=3 8.7->3.0us, down(1536x6912) M=3 53.5->10.3us; full gate 691 passed, test_bf16_gemv_uses_current_stream passes in isolation after GPU contention rerun
- add decode-shape benchmark harness
- add bf16 GEMV CUDA primitive with head-dim generic kernel
- dispatch decode-time linear layers to gemv for M=1
- extend gemv coverage to small decode batches
- Propagate stream closure and stop-sequence termination into scheduler cancellation
- Defer active KV release to the scheduler owner and close metrics safely
- Expose lifecycle counters and cover waiting, active, and allocation-race cleanup
- Return structured finish and error reasons for synchronous generation
- Reject failed online rollout batches instead of training on empty responses
- Verify allocation and extension failures release metrics and KV state
- add astrai/extension/dispatch.py: per-family decision tables over composable Specs with explicit-strict / implicit-loose resolution, ASTR_OPS env overrides, profile presets, and explain traces
- make the axis schema family-owned: register_family takes an axes extractor that snapshots whatever decision axes that family needs from the call, and the core only supplies the axis() predicate vocabulary plus a tensor_axes helper
- drop the central CallContext dataclass; resolve and explain take the raw call arguments, so unregistered handles are probed through supports_call on the same args
- migrate attention and rotary onto family-owned axes with behavior-preserving specs and spec-vs-supports_call mirror tests
- replace the non-ASCII member-of glyph in spec descriptions with plain ASCII " in "
- add a post-step callback hook for checkpoint saves
- preserve updated model, optimizer, and scheduler state
- cover checkpoint ordering with a regression test
- Merge DelayedScaling/DynamicScaling and the abstract FP8Recipe base into one FP8Recipe dataclass with a dynamic flag; dispatch now reads cfg.recipe.dynamic instead of isinstance checks
- Drop the _ActiveOrDefault descriptor and the FP8State property views; the persistent defaults are plain default_* attributes and get_weight_meta takes the active recipe explicitly
- Convert FP8TensorMeta to a NamedTuple of the three per-operand rings
- Update tests to the new API; the autocast context test now asserts _active_config push/restore directly
- the kernel's last block folds amax into the history window and publishes the next scale in-kernel (atomicAdd ticket + fences), replacing the host update chain
- quantize bindings split into quantize(transposed) / quantize_dual with fixed arities and a QuantLayout enum; the python adapter becomes a thin attention-style wrapper over pybind (Optional ring_state at the boundary, no torch.library custom_ops)
- tests: in-kernel fold vs host reference (exact), dual/transposed orientation byte-equality
Benchmark: L20 (sm_89), 1.2B model, full train step. Per-linear fixed overhead 28.8us -> 8.8us; fp8 vs bf16: M=512 77.5ms, M=2048 144.5ms (1.15x), M=8192 527.4ms (1.28x); losses bit-identical.
- serve_runtime.py and train_runtime.py are host-side Docker helpers, so they join train-entrypoint.sh and lib/ under scripts/docker/
- scripts/tools/ now contains only in-container CLIs
- update wrapper call sites, test import, and docker guide references
- serve.sh/train.sh no longer pass --build on up/run; the build subcommand is the only path that rebuilds
- compose services pin image: astrai:latest so run reuses the existing image instead of triggering a rebuild
- runtime parsers leave CUDA_VISIBLE_DEVICES unset for gpu.devices: all; an empty string hid every GPU inside the container
- server service reserves count: all GPUs so CUDA_VISIBLE_DEVICES performs the only filtering, matching the trainer
- wrapper compose() strips an empty host CUDA_VISIBLE_DEVICES before invoking docker compose
- re-enable kFastLoop on the 128x128 CTA for congruous layouts: the base-pair fragment addressing freed the registers the old offset tables spilled, and the predication-free interior loop now wins across the band (fast body 142 SASS instr with zero predicated fallback vs 719/136 generic; 128 regs, no spill)
- move the big/small CTA dispatch boundary from 3/4 to 5/8 wave: with the fast big-CTA loop the crossover sits between 49 and 63 tiles (63-tile rect +8%, 1024^3 now takes the big CTA)
- fuse the linear bias into the GEMM epilogue: FP8Params.bias_ptr adds in fp32 before the single bf16 rounding, replacing the separate out + bias elementwise pass; guarded loads keep N tails exact and batch broadcast falls out of the row-major layout
- resolve Python None bias in the pybind layer (py::object + cast) so ops/fp8.py and fp8.py pass the argument through untouched; drop the _empty_bias sentinel machinery
- add fused-bias tests covering odd N tails, no-bias parity and batched broadcast
Benchmark: L20 (sm_89), CUDA-graph e2e. Big-CTA fast loop + dispatch: 1024^3 102.6->106.3T, 1152^3 128.5->133.3T, 2048^3 173.8->178.2T, 3072^3 180.2->185.3T, 8192^3 196.2->197.7T. Bias fusion (with-bias GEMM vs unfused out + bias): 1024^3 90.5->106.1T (+17%), 2048^3 162.2->178.3T (+10%), 4096^3 178.2->191.1T (+7%). Fused bias differs from the split path by <=1 bf16 ulp and is closer to the fp64 reference. 596 tests pass.
- mm_fp8 accepts 3D operands through the same signature: grid.z slices by batch strides, size-1 batches broadcast (stride 0), inner .t() views fold into the layout tag at zero copy
- fix _LinearFp8 backward crash on 3D [B,L,d] training inputs (flatten before mm_fp8, reduce grad_b over leading dims)
- expose kRasterGroup/kStreamOut as template knobs; drop the 64x128 mid CTA and staged crosswise-B path from dispatch (direct wins everywhere re-measured, including DRAM-streamed B)
- dispatch thresholds grounded in fresh sweeps: m<=64 -> 64x64 CTA (+27% at 64x8192x2048), small-CTA crossover at SM*14/3 total tiles (+13% at 96 tiles), threshold counts batch x per-matrix tiles (+31% at 64x512^3 bmm, +25% at 8x1024x2048)
- remove scripts/tools/bench_fp8_gemm.py (superseded by csrc/tests/fp8_sweep.cu for kernel-level tuning)
Benchmark: NVIDIA L20, E4M3, NT pre-quantized, median of 100-200 iters
- 64x8192x2048: 29.1 -> 22.8 us (94 TF/s)
- 1024x1536x2048: 67.4 -> 59.6 us (108 TF/s)
- bmm 64x512^3: 139.8 -> 106.7 us; bmm 8x1024x2048: 186 TF/s
- regression-free: 4096^3 192 TF/s, 8192^3 200 TF/s, 512^3 unchanged
- split quantize into quantize.cuh, templated on input type (bf16/fp16/fp32)
- rename pybind entry quantize_bf16 to quantize; validate the fmt enum
- fix fp8x2 packing: one 32-bit word packs two pairs (halves were dropped)
- drop the dead OutFp8 template param; GEMM output is always bf16
- fp8_state.reset() restores recipe/format defaults too (test state leak)
- rewrite tests for the two-primitive API with fp32-domain amax references
- g/x/w may each be bf16 or pre-quantized fp8 matching fmt; a pre-quantized operand skips its quantize kernel
- snapshot sx/sw/sg before the ring finalize overwrites the aliased scale slot so the gemm dequantizes with the quantize scale
- forward carries its scale to backward so gradients reuse the forward's scale
- grad_input/grad_weight forced bf16; a pre-quantized g dequantizes before the bias-sum
- regression test: two delayed steps with a changing amax must not leak the scale ratio
- Finalize scale rings inside the quantize kernels: a last-block epilogue (threadfence + counter elect) folds amax into hist, reduces the window and publishes the next scale on device, zero extra launches; _ScaleRing packs [hist | scale | counter] into one CUDA buffer.
- Split FP8QuantizeParams out of FP8Params so each operator owns its fields; linear_forward/backward_fp8 take optional ring arguments.
- Drop the inference weight-quantization cache; the optimizer bumps the weight version every step, so a cache would miss anyway.
- Zero amax scratch via empty + cudaMemsetAsync instead of torch::zeros, cutting a ~50us fill_ dispatch per quantize.
- Stage crosswise-B operands K-major with cp.async (contract >= 8192) and PRMT-transpose per k_seg region in smem, interleaved with the MMAs; the sync LDG + byte-scatter path it replaces was long-scoreboard bound (ncu 4.6 vs 0.4 stalls/issue).
- Load crosswise-A direct with an in-register PRMT transpose; its operands are typically L2-resident and the staging round trip measured as a net loss.
- Enable grouped rasterization for the congruous NT forward (shared B stripe keeps the weight operand hot in L2) and make the smem budget layout-aware (Fp8GemmSmem) while holding two CTAs per SM.
- Annotate ops/fp8.py return types; drop weight-cache and decorator tests, hoist their imports to module level.
e2e 12L/dim1024/B4xT512 fused AdamW: fp8 137.8ms/step vs bf16 210.3ms, 1.53x. Kernel vs cuBLASLt _scaled_mm: fwd 1.03-1.09x, dX 1.33-1.47x, dW 1.30-1.39x (from 1.10/1.42-1.49/1.52-1.56x), before the pre-transposed copies cuBLASLt needs for dX/dW. fp8 train step vs bf16: 1.34x at 2048 tokens (was 1.25x), 1.08x at 512.
- last-block epilogue (threadfence + counter elect) folds amax into hist[idx], reduces the window and publishes the next scale on device — zero extra launches per linear layer
- _ScaleRing packs [hist | scale | counter] into one CUDA buffer; the eager hist-write / max / scale-copy chain and update() are gone
- split FP8QuantizeParams out of FP8Params so each operator owns its fields; linear_forward/backward_fp8 take optional ring arguments
- e2e 12L/dim1024/B4xT512 (fused AdamW): fp8 137.8ms/step vs bf16 210.3ms, 1.53x; fwd 1.82x, bwd 1.50x
- collapse FP8TensorMeta's 12 slots + 6 copy-paste methods into three _ScaleRing objects (hist/idx/scale/initialized + update/seed)
- skip meta allocation entirely on the DynamicScaling path (zero rings, scales measured inline)
- drop write-only FP8State._last_device and unused E4M3_MAX alias
- linear_forward_fp8 accepts pre-quantized w8 (matching fmt) and skips the weight quantize; amax_w returns 0 on that path since no bf16 values are seen
- bias is now fused into the GEMM epilogue for both dtypes, replacing the separate torch-level add (one elementwise kernel per linear removed)
- FP8Params.bias becomes void* with a new bias_scale slot: null scale = raw bf16 bias, non-null = fp8 storage dequantized in the epilogue after the operand scaling and before any output quantization
- ops/fp8.py relaxes the w dtype check to bf16-or-fp8 and passes bias_scale through
- regression test covers w8/b8, w8/bf16-bias and the amax_w = 0 contract vs an explicit quantization reference
- backward used to read the global fp8 flag at loss.backward() time, so calling it outside fp8_autocast silently fell back to bf16 mm (953 ms cublas per step, 49.9% of the model step)
- _LinearFp8(torch.autograd.Function) now owns the fwd/bwd pair: forward captures fmt/recipe/meta on ctx inside the autocast region, backward reads only ctx (scales from the meta rings, masks from ctx.needs_input_grad), so backward is fp8 wherever it runs
- register the aten::linear impl on AutogradCUDA (replaces torch's generated linear formula that calls aten::linear_backward into the bf16 fallback) and keep the CUDA key for inference_mode
- drop the aten::linear_backward override and fp8_linear_backward (dead paths)
- regression test asserts the fp8 backward fires outside the autocast region and grads match the bf16 reference by direction/norm (E5M2 noise)
- model step (0.67B, CE loss, batch 4x1024): backward GEMMs 953 -> 618 ms (1.54x), full step ~1.2x
- add skip_no_fp8 (CUDA + fp8_mm kernel + cc 8.9+) to tests/conftest.py
- use skip_no_cuda / skip_no_kernel / skip_no_fp8 directly in test modules
- drop _GPU alias and tests.extension.conftest re-exports
- remove unused imports (Union in hf_adapter, make_grpo_config in data conftest)
ViperEkura
0z5a