docs: fix stale developer documentation claims

- Move task_alloc/task_free/task_extend/task_cached/task_record_hashes and bind from the PagePool card to a new TaskCacheManager card matching pool.py
- Drop the nonexistent Executor tokenizer attribute and association, add task_cache instead
- Add AllocationStrategy/ContiguousStrategy/PagedStrategy cards and point Allocator/RadixCache composition at PagedStrategy
- Add TaskCacheManager and the allocation strategies to the module overview, add _task_cache to InferenceScheduler
- Fix the design-pattern count in the table of contents (15 -> 16)
- Rewrite the FlashAttnBackend class docstring: packed decode gathers flat K/V via req_to_token and calls flash_attn_varlen_func; dense prefill uses flash_attn_func (no flash_attn_with_kvcache exists)
- Apply the same correction to the backend bullets in internals.md and cuda_kernels.md
- Rename the stale fp8_mma_test.cu reference to fp8_test.cu in cuda_kernels.md

astrai/extension/backend/attention.py

@@ -694,12 +694,13 @@ class CudaBackend(AttentionBackend):
 class FlashAttnBackend(AttentionBackend):
     """FlashAttention backend via the optional ``flash-attn`` package.
 
-    Decode (q_len=1, contiguous cache): uses ``flash_attn_with_kvcache``,
+    Decode (q_len=1, contiguous cache): writes K/V to the pool, gathers
-    which reads K/V directly from the flat pool via cache_batch_idx +
+    flat K/V via the ``req_to_token`` page table, and calls
-    cache_seqlens — no materialized KV gather.
+    ``flash_attn_varlen_func`` over the ragged batch
+    (``qo_indptr``/``kv_indptr``).
 
-    Prefill / non-contiguous decode: falls back to KV gather +
+    Prefill: packed 3-D calls share the ``flash_attn_varlen_func`` path;
-    ``flash_attn_func``.
+    dense 4-D calls go through ``flash_attn_func`` (mask-free only).
     """
 
     @classmethod

docs/developer/architecture.md

@@ -4,7 +4,7 @@
 
 - [Class Diagram](#class-diagram) — Full Mermaid class diagram across 10+ namespaces
 - [Module Overview](#module-overview) — Component inventory per module
-- [Design Patterns](#design-patterns) — 15 documented patterns with classes
+- [Design Patterns](#design-patterns) — 16 documented patterns with classes
 - [Core Relationships](#core-relationships) — 11 key inter-component relationships
 
 ## Class Diagram
@@ -816,8 +816,8 @@ classDiagram
 
         class Executor {
             +AutoModel model
-            +AutoTokenizer tokenizer
             +PagePool kv_cache
+            +TaskCacheManager task_cache
             +InferenceWorkspace _workspace
             +Optional[str] device
             +Optional[torch.dtype] dtype
@@ -845,6 +845,7 @@ classDiagram
 
         class InferenceScheduler {
             +PagePool _cache
+            +TaskCacheManager _task_cache
             +Executor _executor
             +TaskManager _task_mgr
             +Event _stop_event
@@ -888,6 +889,24 @@ classDiagram
             +release(pages)
         }
 
+        class AllocationStrategy {
+            <<abstract>>
+            +alloc(state, prompt_ids) bool
+            +free(state)
+            +extend(state, pos) bool
+            +write_indices(state, prompt_ids)
+            +record_hashes(state, prompt_ids, start_logical_page)
+        }
+
+        class ContiguousStrategy {
+            +write_indices(state, prompt_ids)
+        }
+
+        class PagedStrategy {
+            -Allocator _alloc
+            -RadixCache _prefix
+        }
+
         class KVStorage {
             +int size
             +Tensor k_buffer
@@ -926,14 +945,21 @@ classDiagram
             +bool contiguous
             -KVStorage _storage
             -ReqToTokenPool _req_pool
-            -Allocator _alloc
+            -AllocationStrategy _strategy
-            -RadixCache _prefix
+            +strategy AllocationStrategy
+            +req_pool ReqToTokenPool
+            +bind_tasks(req_indices, seq_lens, workspace, device, start_pos, incremental) KVCache
+        }
+
+        class TaskCacheManager {
+            -PagePool _pool
+            -Dict _states
             +task_alloc(task_id, prompt_ids) bool
             +task_free(task_id)
             +task_extend(task_id, pos) bool
             +task_cached(task_id) int
             +task_record_hashes(task_id, prompt_ids, start_logical_page)
-            +bind_tasks(task_ids, workspace, device, start_pos) KVCache
+            +bind(task_ids, workspace) KVCache
         }
 
     class Task {
@@ -1316,17 +1342,22 @@ classDiagram
     PositionIdStrategy <|-- DocResetPositionId
     PositionIdStrategy <|-- ContinuousPositionId
     StoreWriter <|-- BinWriter
+    AllocationStrategy <|-- ContiguousStrategy
+    AllocationStrategy <|-- PagedStrategy
     RawRollout <|-- RolloutResult
     LaunchStrategy <|-- TorchrunStrategy
     LaunchStrategy <|-- LocalStrategy
     %% --- Composition (strong ownership, part destroyed with whole) ---
     PagePool *-- KVStorage
     PagePool *-- ReqToTokenPool
-    PagePool *-- Allocator
+    PagePool *-- AllocationStrategy
-    PagePool *-- RadixCache
+    PagedStrategy *-- Allocator
+    PagedStrategy *-- RadixCache
+    TaskCacheManager o-- PagePool
     RadixCache *-- RadixNode
     InferenceEngine *-- InferenceScheduler
     InferenceScheduler *-- PagePool
+    InferenceScheduler *-- TaskCacheManager
     InferenceScheduler *-- Executor
     Executor *-- InferenceWorkspace
     InferenceScheduler *-- TaskManager
@@ -1419,7 +1450,7 @@ classDiagram
     Task --> TaskStatus
     InferenceEngine --> AutoModel
     Executor --> AutoModel
-    Executor --> AutoTokenizer
+    Executor --> TaskCacheManager
     TaskManager --> AutoTokenizer
 
 ```
@@ -1436,7 +1467,7 @@ classDiagram
 | **astrai.model** | ModelFactory, AutoModel, AutoRegressiveLM, EmbeddingEncoder, DecoderBlock, GQA, MLA, MLP, DeepSeekMoE, AttnFactory, FFNFactory, RMSNorm, Linear, LoRAConfig, LoRALinear, RotaryEmbedding, Embedding | Neural network model |
 | **astrai.tokenize** | AutoTokenizer, ChatTemplate | Tokenizer and chat template |
 | **astrai.trainer** | Trainer, TrainContext, TrainContextBuilder, BaseStrategy–GRPOStrategy, StrategyFactory, BaseScheduler–WSDScheduler, SchedulerFactory, TrainCallback(Protocol)–MetricCallback, CallbackFactory, RawRollout, RolloutResult, BaseRewardModel, RolloutGenerator, RolloutRunner | Training workflow |
-| **astrai.inference** | InferenceEngine, InferenceScheduler, Executor, InferenceWorkspace, PagePool, KVStorage, ReqToTokenPool, KVCache, Allocator, RadixCache, Task, TaskManager, TaskStatus, StreamDecoder, GenerateResult, BaseSamplingStrategy–SamplingPipeline, FrequencyPenaltyStrategy, ProtocolHandler, ResponseBuilder, OpenAIResponseBuilder, AnthropicResponseBuilder, StopChecker, GenContext, StopInfo, ChatMessage, FunctionDef, ToolDef, ChatCompletionRequest, AnthropicMessage, MessagesRequest, BaseToolParser, ToolParserFactory, SimpleJsonToolParser | Inference service |
+| **astrai.inference** | InferenceEngine, InferenceScheduler, Executor, InferenceWorkspace, PagePool, TaskCacheManager, KVStorage, ReqToTokenPool, KVCache, Allocator, RadixCache, AllocationStrategy, ContiguousStrategy, PagedStrategy, Task, TaskManager, TaskStatus, StreamDecoder, GenerateResult, BaseSamplingStrategy–SamplingPipeline, FrequencyPenaltyStrategy, ProtocolHandler, ResponseBuilder, OpenAIResponseBuilder, AnthropicResponseBuilder, StopChecker, GenContext, StopInfo, ChatMessage, FunctionDef, ToolDef, ChatCompletionRequest, AnthropicMessage, MessagesRequest, BaseToolParser, ToolParserFactory, SimpleJsonToolParser | Inference service |
 | **astrai.extension** | `backend` policy package, `ops` kernel-wrapper package, `fp8.py` FP8 strategy layer, AttentionBackend, TorchNativeBackend, CudaBackend, FlashAttnBackend, attention, attn_backend, ATTN_BACKEND, apply_rotary_emb, is_available | Stable API over attention/rotary/FP8 execution policy and optional CUDA kernels |
 | **astrai.optim** | OptimizerFactory, MuonAdamW, NoraNadamW, ManoAdamW, composite_step/composite_zero_grad/composite_state_dict, partition_optimizer_parameters | Built-in optimizers (`muon_adamw` / `nora_nadamw` / `mano_adamw`) with shared composite-optimizer helpers |
 | **astrai.parallel** | spawn_parallel_fn, setup_parallel, get_rank/get_world_size/get_current_device, only_on_rank, LaunchStrategy, TorchrunStrategy, LocalStrategy, BaseExecutor, ExecutorFactory, NoneExecutor, DDPExecutor, FSDPExecutor, GradientState, AccumOptimizer, AccumScheduler | Distributed parallel & gradient accumulation |
@@ -1478,4 +1509,4 @@ classDiagram
 10. **AutoModel**: `from_pretrained()` loads `config.json` + `model.safetensors`, `_disable_random_init` replaces `nn.init.*` with no-ops
 11. **Protocols**: `OptimizerProtocol` / `SchedulerProtocol` — structural subtyping for `AccumOptimizer` / `AccumScheduler` wrappers
 
-> Document Update Time: 2026-08-22
+> Document Update Time: 2026-08-29

docs/developer/cuda_kernels.md

@@ -305,7 +305,7 @@ cycle belong under `TYPE_CHECKING`.
 
 - **`AttentionBackend`** (ABC): `fwd_decode` / `fwd_prefill` abstract methods, `forward` dispatches by q_len
 - **`CudaBackend`**: CUDA kernel dispatch — decode via `attn_paged_decode` (page_size=1), prefill via `attn_paged_prefill` (ragged batch, `qo_indptr` + `kv_indptr`). Default on GPU.
-- **`FlashAttnBackend`**: Optional flash-attn dispatch with `flash_attn_with_kvcache` fast path.
+- **`FlashAttnBackend`**: Optional flash-attn dispatch via `flash_attn_varlen_func` over gathered flat K/V.
 - **`TorchNativeBackend`**: SDPA with indirect KV cache gather (always-available fallback)
 
 Default priority: cuda > flash > torch. Set ``ASTR_BACKEND=cuda|torch_native|flash``
@@ -415,7 +415,7 @@ nvcc -I csrc -arch=sm_89 -O3 --use_fast_math \
 Test files:
 - `attn_test.cu` — decode + prefill kernels (correctness tables + benchmarks)
 - `attn_paged_test.cu` — paged decode/prefill kernels
-- `fp8_mma_test.cu` — BF16→FP8→BF16 MMA demo (sm_89)
+- `fp8_test.cu` — single-warp bf16→fp8→mma.sync sanity check + full FP8 GEMM correctness (sm_89)
 
 ## Benchmarks
 
@@ -482,4 +482,4 @@ csrc/
 
 Compiled `.so` files are placed in `astrai/extension/lib/`, separate from Python source files.
 
-> Document Update Time: 2026-08-22
+> Document Update Time: 2026-08-29

docs/developer/internals.md

@@ -185,7 +185,7 @@ The extension package separates mechanism from policy:
 Attention computation is decoupled from the model via `AttentionBackend` ABC (`astrai/extension/backend/attention.py`):
 
 - **`CudaBackend`** (default when supported): decode path uses `attn_paged_decode` with `page_size=1` (the `req_to_token` table serves as the page table, each token slot is a single-token "page"); prefill path uses the ragged-batch `attn_paged_prefill` (addresses each request via `qo_indptr` + `kv_indptr` directly against the flat pool).
-- **`FlashAttnBackend`**: optional flash-attn dispatch with `flash_attn_with_kvcache` fast path for contiguous cache; falls back to KV gather + `flash_attn_func`.
+- **`FlashAttnBackend`**: optional flash-attn dispatch; inference paths gather flat K/V from the pool via `req_to_token` and call `flash_attn_varlen_func` over the ragged batch (fp16/bf16 only); dense mask-free training calls use `flash_attn_func`.
 - **`TorchNativeBackend`** (always-available fallback): writes K/V to cache, gathers via `req_to_token` indirect indexing, calls `F.scaled_dot_product_attention`.
 - The `attention(...)` entry point uses cuda > flash > torch priority and chooses another compatible backend when an automatically selected backend cannot handle a call.
 - Resolution precedence is: explicit `attn_backend(...)` context > `ASTR_BACKEND` env > default. An explicit `attn_backend(...)` selection is strict (incompatible calls raise); `ASTR_BACKEND` is a default-level override that falls back to a compatible backend when incapable. Training calls (`fwd=None`, no KV cache) resolve by capability: the CUDA cache kernels cannot run without a cache, so they fall back to flash (mask-free/causal calls only) and finally to torch SDPA.