Yee Man Chan
|
2c8cd844d0
|
added new names for n_experts, n_experts_used and score_func in TextModel and removed their code in KimiLinear in convert_hf_to_gguf.py. Removed unnecessary ggml_cont and GGML_ASSERT in kimi-linear.cpp
|
2026-02-01 08:42:01 +08:00 |
|
Yee Man Chan
|
6216273ede
|
removed all ggml_cont b4 ggml_reshape_4d
|
2026-01-29 08:46:33 +08:00 |
|
Yee Man Chan
|
a6b2c450c8
|
changed hparams.kda_head_dim to hparams.n_embd_head_kda. added TODO comment for class llama_graph_mem_hybrid_k
|
2026-01-29 08:35:35 +08:00 |
|
Yee Man Chan
|
0444a4faa0
|
remove ssm_o_norm_b
|
2026-01-27 13:19:55 +08:00 |
|
Yee Man Chan
|
f1525b3695
|
new class llm_graph_input_mem_hybrid_k to get around the new MLA change. switch the concat order of ggml_concat calls in kimi-linear.cpp to accommodate MLA changes. Removed support for exp_probs_b.weight
|
2026-01-27 11:25:13 +08:00 |
|
Yee Man Chan
|
560190af97
|
fixed find_hparam calls. Fixed e_score_correction_bias to use bias instead of weight. Removed all ssm_conv bias terms.
|
2026-01-21 22:12:21 +08:00 |
|
Yee Man Chan
|
0aea18e718
|
merged dt_bias to SSM_DT. Do -exp(log_A) in convert_hf_to_gguf.py.
|
2026-01-16 12:02:27 +08:00 |
|
Yee Man Chan
|
c163dff4c0
|
sync fork and comment fixing in kimi-linear.cpp
|
2026-01-14 18:01:44 +08:00 |
|
Yee Man Chan
|
2882915258
|
created static function causal_conv1d to abtract similar code for q/k/v
|
2026-01-14 17:26:00 +08:00 |
|
Yee Man Chan
|
18ae7f4684
|
removed unnecessary ggml_cont before ggml_reshape
|
2026-01-14 03:22:53 +08:00 |
|
Yee Man Chan
|
22bc582a82
|
return ggml_tensor * pair in kda_autoregressive and kda_chunking as in ngxson's Qwen3Next improvement
|
2026-01-12 20:32:19 +08:00 |
|
Yee Man Chan
|
59182f5e06
|
fix trailing whitespace
|
2026-01-11 22:06:48 +08:00 |
|
Yee Man Chan
|
93afbedc96
|
moved const llama_model & model; around to follow qwen3next format and see if it cna pass the -Wunused-private-field error
|
2026-01-11 21:44:54 +08:00 |
|
Yee Man Chan
|
6ae66fc40d
|
fix trailing spaces
|
2026-01-11 21:31:35 +08:00 |
|
Yee Man Chan
|
b9360c7fe1
|
MLA KV cache support
|
2026-01-11 15:58:46 +08:00 |
|
Yee Man Chan
|
d26fe50178
|
Moved Aqk computation out of the loop
|
2026-01-10 08:45:57 +08:00 |
|
Yee Man Chan
|
6150bb7b17
|
no clamp version
|
2026-01-09 20:11:45 +08:00 |
|
Yee Man Chan
|
f99913dd5f
|
replaced Akk and Aqk with mul_mat and clamp
|
2026-01-08 13:40:17 +08:00 |
|
Yee Man Chan
|
1099cbf694
|
build_kda_autoregressive is implemented to replace build_kda_recurrent for faster inference. sync'd to b7682
|
2026-01-07 18:42:31 +08:00 |
|
Yee Man Chan
|
e3542ff8a2
|
fixed some comments
|
2026-01-06 11:35:25 +08:00 |
|
Yee Man Chan
|
cfed14e31b
|
naive chunking form implemented
|
2026-01-06 11:23:53 +08:00 |
|
Yee Man Chan
|
aba181ebad
|
removed LOG_INFO
|
2026-01-05 19:21:06 +08:00 |
|
Yee Man Chan
|
66c0c5d8d4
|
Kimi Linear backend agnostic
|
2026-01-05 16:35:19 +08:00 |
|
Yee Man Chan
|
a0269af292
|
removed all hard code
|
2025-12-06 11:51:16 +08:00 |
|
Yee Man Chan
|
9f1265fec1
|
removed some hard coded code
|
2025-12-05 19:51:02 +08:00 |
|
Yee Man Chan
|
27baad43d5
|
kimi linear model implementation
|
2025-12-02 08:35:14 +08:00 |