- add _BindState dataclass and _is_steady_increment() to cache.py
- replace _bind_sig/_bind_seq_lens dual fields with single _bind_state
- replace DecodeSteadyState bare tuple with named dataclass
- use _is_steady_increment() in both PagePool.bind_tasks and Executor.execute_decode
- 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
- 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
- _warmup_cuda_graphs pre-captures graphs at init for power-of-two batch sizes
- Graph enabled only when CudaBackend + supported head_dim + warmup succeeds
- Decode speedups vs no-graph: B=1 2.09x, B=4 1.80x, B=8 1.94x, B=16 1.76x
- 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
- PagedKV::host_kv_len now returns max_context_len instead of max_seq_len
- Eliminates grid-z instability for CUDA graph capture/replay
- Restore skip_no_kernel re-export accidentally removed by ruff --fix
- 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)
- adapt bind_tasks to workspace API and reuse a stable workspace
- drop required checkpoint, randomize default 1B GQA preset
- add config override flag for arbitrary model architectures
- sample the first token from prefill logits without duplicating the prompt tail
- reject incomplete multi-output records before preprocessing alignment
- cover cached generation and partial DPO records with regression tests
- 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
- merge AttentionParams and PagedAttentionParams into one struct
- add attn_kv_source.cuh with ContigKV/PagedKV addressing policies
- template prefill/decode kernels (MMA + scalar) on the KV policy, deleting the four duplicated attn_paged_*.cuh variants
- template dispatcher launchers on KV; single combine kernel
- verify: all correctness tests pass and SASS matches baseline (no perf regression)
- Fix rotary docs to describe cos/sin freqs_cis table, not complex buffer
- Replace attn_prefill with attn_paged_prefill for the CudaBackend path
- Register attn_paged_prefill in kernel overview, layout, and module list
- Add qo_indptr and InferenceWorkspace to architecture class diagram
- Add FrequencyPenaltyStrategy to sampling design patterns
- Extract _step() as the single prefill-group + task_extend + decode primitive
- _run_generation_loop and run_batch now both call it, so the two cannot drift
- run_batch now records prefix hashes (paged mode) and uses input order for
decode, matching the loop thread
- 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
- add InferenceWorkspace with fixed-shape per-step buffers (input_ids, decode mask, KV bind metadata) for CUDA-graph capture
- bind_tasks derives seq_lens from the pool's own _task_len tracking, dropping the seq_lens parameter
- update decode metadata in-place (position_ids, seq_lens, kv_indptr) instead of re-allocating per step
- task_extend advances _task_len in contiguous mode so the pool tracks current length
- skip log_softmax when logprobs are not requested
- SamplingBatchInfo: sample params built once per task set (top_k int32, pinned async H2D)
- position_ids advances by +1 on steady-state decode instead of re-building
- DecodeBindCache: bind_tasks increments seq_lens/kv_indptr, reuses req_pool_indices
- saves ~240us of python/launch overhead per decode step
- Drop hardcoded CACHE_MAX_SEQ=2048 which overflowed at long prompts
- Size prefill pool to prompt_length and decode pool to prompt+5+gen*num_trials
- Unblocks decode/prefill benchmark at prompt 4096+ (was KV cache index OOB)
- compute_num_splits used 2*sm/base, undersplitting at large batch
- single-warp decode blocks host ~11/SM, not 1/2-SM, so B=16 got 3 splits when 8 was optimal
- Grid search on L20: bandwidth saturates near 256-512 total blocks; target 512
- Pass num_passes into base_blocks for the non-paged decode to match the paged path
- B=16 kv=2048: 0.0230->0.0157ms (-32%); paged B=16: 0.0527->0.0243ms (-54%); B=32: 0.0406->0.0241ms (-41%)
- Thread a cudaStream_t through attn dispatchers onto torch's current stream
- Scope the device guard to the entry function so kernels run on tensor device
- DISPATCH_HEAD_DIM now forwards varargs so stream reaches each dispatch
- Parallelize CPU reference kernels with OpenMP (paged test 31s -> 7s)
- Merge decode/prefill standalone tests into attn_test.cu with correctness tables
- Drop bench error column (CPU ref too slow at large sizes)
- Update cuda_kernels.md for the merged test layout
- Converts detached strategy metrics before returning loss output
- Removes redundant item conversion from the trainer loop
- Updates the documented contract and regression tests
- Propagates MoE load-balancing loss through model outputs
- Logs task, auxiliary, and weighted losses across strategies
- Computes only explicitly requested callback metrics
- Preserves tensor compute_loss API and adds regression tests
- 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
- Remove tautological test_trainer assertions that never trained
- Drop grpo isfinite-only smokes and merge frozen-model checks via parametrize
- Merge duplicate tool_parser cases (find/streaming/factory) with parametrize
- Collapse duplicate dataset store/detect_format tests
- Remove misleading scheduler/task tests that asserted the opposite of their names
- Merge signal-handler SIGTERM/SIGINT into one parametrized case
- Drop cross-file grpo strategy duplication kept in online_strategy
- 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
- Add ParallelBuildExtension that dispatches each extension to a subprocess
- 4 extensions compile concurrently (3m34s → 1m1s on L20, ~3.5x faster)
- Default 8 workers, override with BUILD_PARALLEL=N
- ManoAdamW now uses torch.optim.AdamW(fused=True, betas=(0.9, 0.95)) matching MuonAdamW, eliminating a confounding variable in optimizer comparison experiments
- only NoraNAdamW retains NAdamW, which is correct per the Nora paper design
- add GradSNRTracker to metric_util.py computing SNR = E[g]^2 / Var(g) via per-parameter EMA moments
- add grad_snr_tracker field to TrainContext (instantiated by default)
- register grad_snr in MetricCallback, update tracker on each optimizer step before metrics are recorded
- add grad_snr to default --metrics in train.py CLI
- implement Mano (v2) with axis-rotating tangent projection and manifold normalization, replacing Newton-Schulz iteration
- composite ManoAdamW reuses partition_optimizer_parameters and composite helpers
- register mano_adamw in OptimizerFactory, export Mano and ManoAdamW
- add --mano_momentum and --mano_nesterov CLI options in Optimizer group
- add mano_adamw hyperparameters branch in train.py
- document mano_adamw in params.md
- add tests for single-step projection, axis alternation, factory registration, closure, and resume
- add GroupedOption/GroupedCommand (no third-party dep) that tags each option with a group label and renders help in labeled sections
- add opt() shorthand wrapping click.option with cls=GroupedOption
- tag all ~55 options into 10 groups aligned with params.md chapters
- add astrai/optim/composite.py with shared step/zero_grad/state_dict/param_groups helpers and OptimizerFactory
- rename MuonMix to MuonAdamW (matches registered name muon_adamw) and file to muon_adamw.py
- use @OptimizerFactory.register decorator in each optimizer module instead of post-import registration in __init__
- fix closure being invoked once per sub-optimizer in MuonAdamW.step (now exactly once via composite_step)
- NoraNAdamW.step now forwards closure correctly
- 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
- apply_rotary_emb now checks torch.is_grad_enabled() before dispatching to CUDA kernel
- Training (grad enabled) uses torch complex multiply path which supports autograd backward
- Inference (inference_mode/no_grad) uses CUDA kernel as before
- Without this fix, training backward would crash — the CUDA kernel has no autograd backward()
- CUDAExtension module names changed from astrai.extension.<name> to astrai.extension.lib.<name>
- Compiled .so files now land in astrai/extension/lib/ instead of alongside Python source
- loader.py imports from .lib.<name> subpackage
- Add astrai/extension/lib/__init__.py to make lib a proper package
- Separates compiled artifacts from Python source for cleaner directory structure
- Precompute page_table and decode_mask on KVCache once per step in PagePool.bind_tasks, instead of per-layer in CudaBackend/TorchNativeBackend
- Skip frequency penalty history tensor construction when all penalties are 0 in Executor.execute_decode
- Omit FrequencyPenaltyStrategy from sampling pipeline when penalty is 0
- Deduplicate get_active_tasks calls in scheduler loop (3 to 1), remove redundant sorted() on decode tasks
- Benchmark (L20, bf16, CUDA backend): B=1 9.48->9.40ms (+1%), B=4 10.73->9.89ms (+8.6%), B=8 10.77->10.13ms (+6.4%)
- Add cu128/cu130 build matrix to release workflow
- Parametrize Dockerfile and docker-compose with CUDA_TAG build arg
- Allow csrc/ and setup.py in docker context via .dockerignore
- Add nvcc/torch CUDA version mismatch preflight warning in setup.py
- Add cuda_toolkit_version() helper in csrc/build.py
- Use at::IntArrayRef explicitly to fix ATen overload ambiguity
- Guard kernels with CUDART_VERSION >= 11020 check
- Remove invalid [tool.pip] section from pyproject.toml
- Add broadcast_state_dict to sync state_dict from rank-0 to all ranks
- Fix create_ref_model returning None on non-rank-0 under FSDP
- Fix sync_old_model only updating old_model on rank-0 under FSDP
- Split skip_no_cuda/skip_no_kernel markers and hoist to top-level conftest
- Add distributed tests for broadcast_state_dict and create_ref_model
- Replace per-tile page lookup with per-element lookup in load_tile
- Remove page_ok gate and scalar fallback in launch_paged_decode_mma
- Unified path works for any page_size (L1-cached when page_size >= BC)
- HBM BW: 12% → 73%, decode throughput: 2,250 → 2,606 tok/s (B=32)
- Scales to 5,232 tok/s at B=128 (2.54x vs torch native)
- CudaBackend: paged decode via attn_paged_decode, prefill via attn_prefill
- Decode uses req_to_token as page_table with page_size=1
- Falls back to TorchNativeBackend when kernel unavailable
- Rename forward_decode/forward_extend to fwd_decode/fwd_prefill
- Register ATTN_BACKEND.CUDA in _BACKEND_REGISTRY
- Replace all buffer[layer_id][loc] double indexing with buffer[layer_id, loc] single advanced indexing in cache.py and attention.py
- Revert KVStorage buffers back to 4D [n_layers, size, n_kv_heads, head_dim], remove leftover 3D reshape/view in MLA path
- Update docs/guides/inference.md, docs/developer/internals.md, docs/developer/architecture.md to reflect new PagePool/KVStorage/ReqToTokenPool/KVCache classes
- Extract _resolve_base_type and _validate_component as module-level helpers
- Replace ForwardRef._evaluate private API with eval in module namespace
- Remove broad except Exception in __init_subclass__, _component_base always set
- Replace direct _entries mutation in strategy.py with register() call form
- Remove dead TOKENIZER_CLASSES registry from AutoTokenizer
- Extract ModelFactory(BaseFactory[nn.Module]) as pure factory
- AutoModel now inherits only nn.Module, no factory state
- Move @AutoModel.register to @ModelFactory.register in transformer.py and encoder.py
- Replace hand-rolled BaseConfig (from_dict/to_dict/_coerce/_unwrap_optional) with pydantic.dataclasses
- from_dict now uses cls(**d), to_dict uses dataclasses.asdict + json.dumps filter
- TrainConfig: required fields are now truly required (no default=None), delete manual validate()/__post_init__
- Remove dead required() helper and metadata={'help': ...} annotations
- Fix gradient_checkpointing_modules type from List[str] to List[type]
- Add pydantic>=2.0 as direct dependency in pyproject.toml
- Add numpy-style Parameters docstrings to all config classes
- Enable use_attribute_docstrings in BaseConfig for schema generation
- LoRAConfig also migrated to pydantic dataclass
- Move create_ref_model from astrai/parallel/utils.py into executor.py
- Remove unused ColumnParallelLinear/RowParallelLinear (module.py)
- Update imports in strategy.py and train_context.py
- Drop unused astrai.parallel.utils and astrai.parallel.module
- unshard() and full_tensor() are collective ops, all ranks must participate
- Old code returned None on non-rank-0 before calling unshard, causing deadlock
- Fix: all ranks unshard/full_tensor, only rank-0 keeps the result
- Move create_ref_model to parallel/utils.py, accept executor+model directly
- Guard create_ref_model and sync_old_model against None on non-rank-0
- FSDP params are DTensors sharded across ranks
- torch.nn.utils.clip_grad_norm_ computes LOCAL norm only
- Each rank would clip by a different factor, causing gradient divergence
- Fix: compute local norm, all-reduce squared sum, sqrt for global norm
- Default reshard_after_forward=False (forward then backward makes reshard redundant)
- Reduces per-step time by ~19% (1033ms to 839ms on 2xL20)
- Add __getstate__/__setstate__ to drop cached _compiled Jinja2 template
- Jinja2 Template.root_render_func is a dynamic closure unpicklable by reference
- cached_property rebuilds the template lazily on first render after unpickle
- Remove H5Store, H5Writer, save_h5/load_h5 and h5py dependency
- MmapStore (bin) is the sole pre-tokenized storage backend
- Move setup_logging after imports to fix E402 in __init__.py
- Clean up unused imports across test files
- Move inline test imports to file top
- Remove --max_tokens, let scheduler use max_seq_len - prompt_len
- Rename --cache_len to --max_seq_len to match engine naming
- Unify sampling defaults to 0.8/50/0.95
- Add EOS token at end of response in both conditional and unconditional passes so model also predicts when response should end
- New --append_eos/--no-append_eos CLI flag (default: enabled) with graceful fallback when tokenizer has no EOS
- Replace hand-rolled StreamDecoder (O(n^2) full-history re-decode per token) with tokenizers.decoders.DecodeStream
- Keep O(1) bounded token buffer internally via prefix drain instead of accumulating all token IDs
- Simplify flush_remaining to no-op since stream always emits completed text per step
- Benchmark on 8000 tokens: 2305ms -> 3.9ms (~592x speedup)
ViperEkura