V0.0.1.0209#28

Merged

luozixin2 merged 80 commits intoSJTU-DENG-Lab:v0.0.1.0209from

luozixin2:v0.0.1.0209

Feb 9, 2026

Commits on Dec 15, 2025

feat(kv-cache): add dtype protocol utilities for KV cache
lzx
committed
feat(kv-cache): add multi-dtype support (BF16/FP16/FP32/FP8) to KV cache kernels
lzx
committed

Commits on Dec 22, 2025

fix(test): fix import path in test_dllm_kv_cache_store
luozixin2
committed
test(kv-cache): add FP8 roundtrip tests for unified and distinct layouts
luozixin2
committed
fix(attention): optimize shared memory usage in prefix_prefill kernel
luozixin2
committed

Commits on Dec 24, 2025

Commits on Dec 25, 2025

Commits on Dec 27, 2025

fix: dllm_flash_attn_decode_kernel recompilation problem fixed
drewjin
committed

Commits on Dec 28, 2025

Commits on Dec 29, 2025

Commits on Dec 30, 2025

Commits on Dec 31, 2025

Commits on Jan 1, 2026

Improve W8A8/W4A8 quality by using FP16 scales instead of BF16
luozixin2
committed

Commits on Jan 5, 2026

Commits on Jan 6, 2026

Commits on Jan 12, 2026

Commits on Jan 13, 2026

fix: 修复 FP8 KV cache RunningMax 策略中的 scale 更新逻辑
luozixin2
committed
chore: 移除 .cursor 目录并添加到 .gitignore
luozixin2
committed
Merge commit '67686e0' into feat/kv-cache-fp8-support
luozixin2
committed

Commits on Jan 14, 2026

feat: optimize W8A16 decode and FP8 KV varlen path
luozixin2
committed

Commits on Jan 16, 2026

Commits on Jan 18, 2026

feat: 支持 GPTQ Marlin 和 AWQ Marlin 量化格式
luozixin2
committed
chore: 从仓库移除 benchmark_results
luozixin2
committed
升级 quantize_model.py 为真正的 GPTQ/AWQ 量化路径
luozixin2
committed
refactor: 优化代码结构和消除重复逻辑
luozixin2
committed

Commits on Jan 24, 2026

fix: 修正 bench 中 prefill/decode 吞吐量的平均值计算
luozixin2
committed

Commits on Jan 25, 2026

Commits on Jan 26, 2026

perf: cache linear forward dispatch for CUDA Graph
luozixin2
committed

Commits on Jan 27, 2026

Fix static+CUDA Graph mode and add benchmark configs
luozixin2
committed

Commits on Jan 28, 2026

Commits on Feb 9, 2026