12 commits

Author	SHA1	Message	Date
Yuhao Tsui	cea07d1998	Feat: Clear cache during weight loading to prevent OOM on GPUs with <=8GB VRAM ... This change explicitly clears CUDA cache during weight loading to mitigate memory fragmentation issues, particularly beneficial for low-VRAM GPUs.	2025-02-24 10:09:42 +08:00
Atream	e8e02e5ccc	support Moonlight	2025-02-23 14:21:18 +00:00
Atream	5ec33d046d	optimize gguf dequant, save mem, support Q2_K ... use marlin for lm_head, lm_head only calc last token for prefill extend context window to 19K for DeepSeek-V3/R1 within 24GB VRAM	2025-02-22 06:13:01 +00:00
Atream	038bc30888	fix precision bug imported by position_ids in 0.2.0	2025-02-17 09:23:14 +00:00
Atream	1946493f2d	warm_up before capture	2025-02-14 15:52:21 +00:00
Atream	bb35dc5b0d	init support for MLA using Attention kernel	2025-02-13 15:01:14 +00:00
liam	d07087a7e2	⚡ support R1 force thinking	2025-02-11 15:43:41 +08:00
yangshen	ee72cee050	Fix: the tokens return by prefill_and_generate	2024-09-05 05:29:23 +00:00
chenxl	4d1d561d28	[feature] release 0.1.3	2024-08-28 16:11:43 +00:00
Atream	412055d450	[feature] experts can be injected using CPUInfer ... [fix] fix ktransformers interface when use new CUDAGraphRunner [fix] fix YAML and optimize logic, the top rule has the highest priority	2024-08-14 16:10:54 +08:00
chenxl	f5f79f5c0e	[ADD] support multi-gpu qlen>1 q5_k	2024-08-12 11:41:26 +00:00
chenxl	18c42e67df	Initial commit	2024-07-27 16:06:58 +08:00