- parameterize warp tile (WarpM/WarpN) in Fp8GemmTraits; MMA loops, fragment arrays and epilogue scale with kMt/kNt instead of the fixed 64x32/4x4, enabling cuBLAS-style 64x64 CTAs of 32x32 warps
- dispatch by output tiling (grid-searched via csrc/tests/fp8_sweep.cu): fewer than 48 output tiles take 64x64/32x32 with a lean ring (4 CTAs/SM fill the wave-quantization gap: 512^3 goes 16 -> 64 CTAs); larger shapes keep 128x128 with the kStages+1 ring
- kStages+1 canonic ring rotation drops the post-compute barrier on the congruous path (one __syncthreads per k-tile); LeanRing keeps the kStages ring for the small CTA; direct-crosswise operands always rotate kStages+1 (their prefetch issues right after barrier 1 and would race a lean ring - caught by the pure C layout suite)
- stage the bf16 epilogue through the reclaimed operand smem: swizzled scatter + barrier + coalesced 16B copy-out replaces 8 disjoint 16B per-warp segments (~50% write efficiency before)
- hoist per-lane ldmatrix swizzle offsets out of the mainloop (stage-relative table + ring-base add) so the innermost loop stops recomputing IMAD/LOP3 address chains
- bypass the torch.library dispatch for real CUDA tensors in quantize/mm_fp8 wrappers (~5us/call, ~40% of a 512-wide call's wall time); fake/subclass tensors keep the custom_op route
vs the previous kernel + python path, wall clock on NT squares: 512^3 52 -> 13us (4.0x, 5.2 -> 20.5 TF, now 1.36x cuBLAS _scaled_mm), 1024^3 1.05x, 2048^3 1.02x (46.9 -> 48.2 TF kernel-only); correctness: 4 layouts x 6 shapes pure C suite PASS, 588 pytest PASS
- 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
- loader.py: lazy/cached import; is_available defers the actual load; get_module raises on unavailable
- ops/{attention,rotary,fp8}: use get_module instead of touching private _modules or their own _mod() cache
- package-data: ship astrai.extension.lib *.so in built wheels (non-editable installs previously lost every kernel)
- route dX/dW through the fused 128x64 fast kernel via contiguous transposes
- drop the legacy 64x64 kernel, cutting dX 1.55->0.38 ms and dW 1.28->0.26 ms
- sync all threads after cp.async.wait_group to fix sporadic NaN in large GEMMs
- add fp8_mm_prequant_fp8 custom op for FP8-in/FP8-out GEMM
- keep training attention on dense 4d tensors
- use packed 3d tensors with KV cache for inference
- extend CUDA rotary embedding to packed 3d inputs
- adapt torch, CUDA and FlashAttention backend dispatch
- store page-table, request-row, and cache-location indices as int32
- preserve CUDA graph replay with bit-exact logits and KV cache coverage
- improve B=1 decode latency by 1-6% across 1K-32K contexts on L20
- amax for delayed scale was the quantized max (always ~448), so scale collapsed to 1
- this made fp8 gradients diverge (cosine 0.05) and training stall
- stop w/x transpose-quantize amax from polluting the grad scale
- fp8_ops is the only module touching the pybind (kernel interface)
- fp8.py keeps scaling state, delayed amax and aten::linear dispatch
- remove circular imports between old fp8_ops/fp8_state/fp8_dispatch
- cast gradients and inputs to weight.dtype instead of hardcoded bf16
- single code path covers bf16 and fp32 models, no branch needed
- gradient dtype now matches the leaf parameter dtype exactly
- rename output pointer field o to o_ptr for consistency with q_ptr/k_ptr/v_ptr
- regroup AttentionParams fields by responsibility and fix misleading comments
- drop unused max_seq_len/total_q fields and paged decode max_seq_len arg
- drop redundant group_size param from decode launchers (computed from p)
- Eliminate core/ directory into cache/, runtime/, network/ subpackages plus flat modules
- Split cache.py (647 lines) into cache/{buffer,strategy,pool}.py by layer
- Add explicit ContiguousStrategy, make AllocationStrategy a real ABC
- Move TaskCacheState to cache/strategy.py, drop string forward references
- Rename api/ to network/, server.py to app.py
- Move sample.py into runtime/ alongside executor and graph
- Simplify TaskCacheManager.__init__ to single pool param
- Expose pool.strategy and pool.req_pool as public properties
- Fix KVCache import in attention_backend.py (TYPE_CHECKING guard)
- Fix steady-state decode reading uninitialized position_ids on first step
- merge _generate_streaming/_generate_non_streaming into single _generate() with stream flag
- delete dead GenerationRequest class and generate_with_request method
- inline _next_token helper into generate_async
- replace flash-attn double-checked locking with functools.lru_cache
- extract _write_and_gather_kv helper shared by TorchNative/FlashAttn backends
- inline _kv_cache_is_contiguous into its sole call site in FlashAttnBackend
- change default backend priority from flash>cuda>torch to cuda>flash>torch
- add ASTR_BACKEND env var to override default backend at resolve time
- add supports_graph() static method to AttentionBackend ABC, override in CudaBackend
- replace isinstance(get_backend(), CudaBackend) with get_backend().supports_graph() in executor
- add torch.cuda.is_available() guard to CudaBackend.supports()
- Pre-allocate decode_out in InferenceWorkspace so attn_paged_decode does not call torch::empty inside graph capture
- Wire decode_out through KVCache, PagePool.bind_tasks, and CudaBackend.fwd_decode
- Run live forward before graph capture to get valid output (graph pool memory is zeroed after capture block exits)
- Greedy generation with graph replay is bit-exact across all batch sizes
- Decode speedups vs no-graph: B=1 2.09x, B=4 1.80x, B=8 1.94x, B=16 1.76x
- Decode with contiguous cache uses flash_attn_with_kvcache instead of materializing full KV via gather + flash_attn_func
- _backend_supports allows FlashAttnBackend for decode (q_len==1) even with explicit mask
- Decode speedups vs TorchNative (B=1,4,8,16 mean): cuda 1.55x, flash 1.40x, torch_native 1.00x
- Read K/V directly from flat pool via cache_batch_idx + cache_seqlens, zero-copy view reshape
- Remove unused local variable b in attention_backend.py
- Remove unused variable rank0_sd in test_broadcast_state_dict.py
- Remove unused imports across test files
- New CudaGraphContext class: warmup -> capture -> replay lifecycle
- One graph per batch_size key, all inputs at fixed workspace addresses
- Added position_ids buffer to InferenceWorkspace (required for graph capture)
- Graph only activates when CUDA backend is the current backend
- Default off (opt-in) due to slight numerical divergence in graph replay
- Sampling stays outside the graph (torch.multinomial uses mutable RNG)
- Resolved circular import: KVCache -> TYPE_CHECKING in attention_backend.py
- Replace per-.cu-file static cached tensors with workspace-managed pre-allocated buffers
- InferenceWorkspace now owns decode_o_part / decode_ml_part (mirrors FlashInfer's workspace pattern)
- KVCache carries the buffers through the backend -> C++ kernel chain
- C++ kernels accept optional pre-allocated buffers; fallback to alloc_split_partials for backward compat
- Pre-allocates once at Executor init, zero allocation in the decode hot loop
- Prerequisite for CUDA-graph capture (all kernel addresses are stable)
- Preload V into shared memory alongside K to eliminate per-element KV address lookups in the inner softmax/accum loop (doubles smem)
- Cache split-KV partial tensors (o_part, ml_part) with static tensors instead of per-call allocation in both decode and paged-decode paths
- Force is_causal=True in CUDA decode backend (decode is always causal)
- add FlashAttnBackend (ATTN_BACKEND.FLASH) using flash_attn_func with KV-cache gather + GQA, mirroring TorchNativeBackend
- add flash_attn_available() probe gated on compute capability plus a real-kernel smoke test, cached at first use
- lazy-import flash-attn via importlib so it stays an optional dependency, raising clear errors when unusable
- add 'flash' optional extra (flash-attn>=2.6) and export the new backend
- hoist prefill qo_indptr into the workspace so CudaBackend.fwd_prefill does not rebuild it per layer
- cache has_freq in SamplingBatchInfo to drop the per-step GPU any() sync
- drop pin_memory host staging for input_ids; sync copy suffices for a small batch
- bind_tasks builds kv_indptr (prefix sum of seq_lens) a single time
- fwd_decode/fwd_prefill reuse it instead of rebuilding per layer
- Removes 24 cumsum launches per decode step (was ~1ms/step at B=4)
- Decode B=4: 9.60 -> 7.82 ms/step (-18.5%), +22.8% tok/s
- CudaBackend.fwd_decode passes attn_mask directly instead of kv_cache.decode_mask
- TorchNativeBackend derives pos_mask from attn_mask[:,0,0] on decode
- Drop decode_mask and page_table fields from KVCache and bind_tasks
- alloc_split_partials now uses torch::empty: the split kernel writes every slot it owns, so the per-call zeros/full memset was pure overhead (2 kernels per layer per step)
- decode split-KV MMA kernels now run a true multi-stage cp.async pipeline (wait_group<STAGES-1> instead of wait_group<0>), keeping STAGES-1 tile loads in flight; the old wait_group<0> serialized load and compute so deeper STAGES made no difference
- add a fallback path when ntiles < STAGES to avoid a race on the last tile
ViperEkura