ViperEkura/AstrAI
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
646 Commits 1 Branch 14 Tags
925cbedc93bd1a637ac2399ffbe53b9a97bb2a10
Commit Graph
15 Commits
Author SHA1 Message Date
ViperEkura 4b25664c79 perf: precompute kv_indptr once per decode step 
- 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
 2026-08-01 16:09:26 +08:00
ViperEkura 91acaf4b0b refactor: unify attention mask to single attn_mask tensor 
- 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
 2026-08-01 15:49:26 +08:00
ViperEkura 50cfd0d555 perf: reduce decode overhead in scheduler and executor 
- 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%)
 2026-07-31 14:50:16 +08:00
ViperEkura 3067a8e1a6 feat: unify attention backend with multi-dim mask support 
- Add attention() functional entry delegating to active backend
- GQA/MLA forward calls attention() instead of inline cache/SDPA
- CUDA kernels support 2D/3D/4D mask via mask_h_stride field
- CudaBackend.fwd_decode builds 2D padding mask for mixed seq_lens
- KVCache.max_len precomputed in bind_tasks to avoid GPU sync
- batch==1 decode short-circuits mask=None
- Split tests into conftest, test_backend, test_backend_equivalence, test_kernel_mask
- 440 tests pass, L20 decode 1.44-1.60x speedup vs torch native
 2026-07-30 20:38:34 +08:00
ViperEkura df979b4469 refactor: use single-index access and update docs for cache architecture 
- 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
 2026-07-30 17:47:04 +08:00
ViperEkura deb2d7e127 refactor: rebuild KV cache with three-layer separation architecture 
- Replace CacheView/ContiguousCache/PageCache with SGLang-inspired design: KVStorage (flat token-level NHD buffers [n_layers, size, H, D]), ReqToTokenPool (index table [req_idx, pos] -> token_slot), Allocator + PrefixCache (slot allocation with LRU and prefix sharing)
- Add KVCache as pure dataclass passed to model: k_buffer, v_buffer, req_to_token, req_pool_indices, seq_lens, out_cache_loc
- PagePool orchestrates all three layers, supports contiguous mode (pre-allocated per-request blocks, default) and paged mode (page_size=1 or >1 with dynamic allocation and prefix caching)
- Attention layers now do raw buffer indexing instead of opaque write/gather method calls on CacheView objects
- Update executor.bind_tasks signature: seq_lens list + start_pos
- Rename paged_cache -> kv_cache throughout model/ and inference/
 2026-07-30 17:19:06 +08:00
ViperEkura ccf728a1b7 perf: eliminate GPU syncs in contiguous cache write/gather hot paths 
- Replace .tolist() calls with _total_len in gather(); move _slot_len updates from per-layer write to once-per-step bind_tasks

- Use torch.as_tensor instead of torch.tensor in decode penalty history construction
 2026-07-21 16:40:56 +08:00
ViperEkura a24a7b4da5 perf: merge decode batch for 10x throughput 
- merge all active decode tasks into single forward pass (was grouped by next_pos)
- add per-task write_positions to ContiguousCacheView for correct KV writes
- override ContiguousCache.task_cached (base returned 0, caused prefill loops)
- add --cache_len/--frequency_penalty/--rep_window to generate.py
- chunked batch processing with tqdm progress

bench (1.2B model, 128 prompts, 64 tok, batch=128):
  before: 77.2s, ~111 tok/s
  after:   7.1s, ~1210 tok/s (10.9x)
 2026-07-18 08:50:46 +08:00
ViperEkura 849e1e00a3 refactor: clean up inference design patterns 
1. KVCache base: add default task_cached/task_record_hashes, remove getattr from scheduler
2. Remove page_size param from scheduler constructor (ContiguousCache-only)
3. InferenceEngine expose cache param for KVCache injection
4. Rename page_cache -> kv_cache in Executor
5. Move stream_callback from Task to TaskManager._callbacks dict
6. TaskManager.clear_queues clears callbacks
 2026-07-05 11:41:54 +08:00
ViperEkura 5416c2e8fb perf: replace paged KV cache with contiguous ContiguousCache, decode all groups 
- Add KVCache/CacheView abstract base classes in cache.py
- Add ContiguousCache (contiguous per-slot buffer, default) alongside PageCache (paged, renamed from old KVCache)
- Merge make_table_tensor + bind into bind_tasks on KVCache interface
- Remove task_cached/task_record_hashes from base class (PageCache-only)
- Scheduler: decode all position groups instead of just the largest (eliminates 63% group skip rate)
- Scheduler: accept optional cache param for swapping implementations
- Model layer type hints use CacheView base class
- Batch 1-32: 1-7% speedup from eliminating Storage.gather overhead
- All 183 inference tests pass
 2026-07-05 11:34:36 +08:00
ViperEkura 17d6eaa2f2 refactor: rewrite humaneval evaluation with functional pipeline design 
- fix KeyError race condition in inference cache touch()
- EvalConfig dataclass for centralized configuration
- load->generate->extract->test->score->report pipeline
- two-phase generation+testing for max GPU utilization
- signal-based SIGALRM timeout protection for code exec
- suppress subprocess stdout/stderr pollution
 2026-07-05 07:58:28 +08:00
ViperEkura cb8dcb97ea refactor : 移除 -> None 返回值标注，拆分 FSDP 参数，新增 mmap 数据集存储 
- 删除所有 def 函数 -> None 返回值类型标注
- FSDPExecutor 参数从 **kwargs 拆为显式声明，None 值自动过滤
- 新增 MmapStorage (bin) 存储后端，基于 numpy.memmap 零拷贝加载
- 新增 save_bin/load_bin/json_to_bin 工具函数
- detect_format 支持 bin 格式自动检测
 2026-05-28 13:57:06 +08:00
ViperEkura e3382f6bb5 fix: 修复推理引擎 batch decode 中多项正确性与并发问题 
- scheduler: decode 分组由幂次分桶改为精确 next_pos，消除 KV cache 位置错乱
- task: activate() 加锁操作 active_tasks，消除数据竞争
- engine: wait_completion 加超时，防止分配失败时永久死锁
- sample: TopKStrategy 向量化为 per-sample threshold，尊重各 task 的 top_k
- cache: Storage.write/gather 中 -1 页改用 mask 处理，防数据污染
- executor: prefill 逐 task 循环改为单次 tensor 调用
 2026-05-14 21:31:39 +08:00
ViperEkura 205b40bd28 refactor: 重构 cache 和 inference 参数体系，分离存储与分配 
- 合并 GenerationRequest/GenerationParams，统一 max_tokens 参数名
- PagePool/PrefixCache 分离为 Allocator + PrefixCache + PagePool
- 拆分 KV 存储为独立 Storage 类，PagedCache → KVCache，CacheView → KvcacheView
- Allocator.inc_ref 移除 LRU 防止竞争，Storage.write 增加负页防御
- Allocator/PrefixCache/TaskTable 加 threading.Lock 保证线程安全
- server.py uvicorn.run 改为传 app 对象修复导入错误
- benchmark.py 适配 KVCache 新 API
 2026-05-14 20:05:08 +08:00
ViperEkura 2196c34c52 refactor: 重构 inference 模块架构，引入设计模式并分组文件 
- 新增 protocol.py 协议层，Template Method 模式消除流/非流分支 45% 重复
- SSEBuilder 统一 SSE 构造，StopChecker 独立 stop_sequence 检测
- AnthropicHandler 追踪已产出文本，修复 stop 时重复 delta
- server.py 路由从约 100 行缩减至 3 行
- 拆分为 core/（cache/executor/scheduler/task）和 api/（protocol/server）
- 外部保持二级导入路径（from astrai.inference import Name）
- 删除所有分隔线注释，代码按语义自然分组
 2026-05-14 17:42:37 +08:00