| .. |
|
1 - Offload Bitnet token embeddings to the GPU.md
|
…
|
|
|
2 - Offload Bitnet token embeddings to the GPU - the right way.md
|
…
|
|
|
3 - Merge mainline llama.cpp.md
|
…
|
|
|
4 - Simdify and multi-thread tanh.md
|
…
|
|
|
5 - Fusing a mat mul op followed by a scale op on the CPU.md
|
…
|
|
|
6 - IQ4_K_ SOTA 4-bit quantization.md
|
…
|
|
|
7 - Adding IQ2_K_ IQ3_K and IQ5_K.md
|
…
|
|
|
9 - Fused soft cap and SIMD-ified GeLU.md
|
…
|
|
|
10 - iq4_k_ speedup quantization by a factor of _2.md
|
…
|
|
|
11 - Faster iq3_k and iq5_k quantization.md
|
…
|
|
|
12 - q2_K_ allow it to detect ternary nets and quantize accordingly.md
|
…
|
|
|
13 - Adding IQ2_TN for use with ternary models.md
|
…
|
|
|
14 - Adding IQ6_K.md
|
…
|
|
|
16 - Fix Makefile.md
|
…
|
|
|
17 - Merge mainline - Aug 12 2024.md
|
…
|
|
|
19 - Skip barriers of noops.md
|
…
|
|
|
20 - iq2_k_ slightly better bpw - accuracy compromise.md
|
…
|
|
|
21 - quantize_stats_ print rmse and max error as fraction of _x_.md
|
…
|
|
|
22 - AVX2 quantization for Q8_K.md
|
…
|
|
|
23 - iq4_k tweak.md
|
…
|
|
|
24 - softcap_ minor improvement.md
|
…
|
|
|
27 - Faster Gemma2.md
|
…
|
|
|
28 - Binary KQ mask.md
|
…
|
|
|
31 - Fix build when iqk_mul_mat is disabled.md
|
…
|
|
|
32 - Zen4 Flash Attention.md
|
…
|
|
|
33 - Do not process prompts containing binary data for escapes.md
|
…
|
|
|
35 - Fix Zen4 Flash Attention.md
|
…
|
|
|
36 - Zen4 Flash Attnetion 2.md
|
…
|
|
|
37 - Performance improvements for legacy quants on ARM_NEON.md
|
…
|
|
|
38 - Zen4 Flash Attention - bf16 support.md
|
…
|
|
|
39 - Add support for bf16 to iqk_mul_mat.md
|
…
|
|
|
40 - Adding bf16 support to CUDA.md
|
…
|
|
|
41 - iqk_mul_mat_ARM_NEON_ adding bf16 support.md
|
…
|
|
|
42 - Adding fused rms_norm.md
|
…
|
|
|
43 - iq2_tn_ slightly faster PP on Zen4.md
|
…
|
|
|
44 - Adding IQ1_TN - 1.6875 bpw for TriLM ternary models.md
|
…
|
|
|
45 - Add CUDA support for IQ1_TN.md
|
…
|
|
|
46 - IQ1_TN Metal implementation.md
|
…
|
|
|
47 - iq2_tn_ slightly better performance on AVX2.md
|
…
|
|
|
48 - AVX2 Flash Attention.md
|
…
|
|
|
49 - ARM_NEON Flash Attention.md
|
…
|
|
|
50 - AVX2 Flash Attention 2.md
|
…
|
|
|
51 - Quantized Flash Attention for all supported CPU platforms.md
|
…
|
|
|
52 - Fix bug and D _ 128 case for Q8_0 k-cache.md
|
…
|
|
|
53 - Quantization mixes tweaks.md
|
…
|
|
|
54 - Improve Q4_0 and Q8_0 performance on AVX2_Zen4.md
|
…
|
|
|
55 - Improve Q5_0 performance on AVX2.md
|
…
|
|
|
56 - BF16 support on Metal.md
|
…
|
|
|
57 - AVX2_Zen4 horizontal sums.md
|
…
|
|
|
58 - Fix compiler warnings.md
|
…
|
|
|
61 - Adding ability to have meta data per tensor row.md
|
…
|
|
|
62 - Use fp32 for K_Q in Metal FA implementation.md
|
…
|
|
|
64 - Better sub-3-bit quantization mixes with a qkv tensor.md
|
…
|
|
|
65 - Adding SWIGLU unary op.md
|
…
|
|
|
66 - CUDA non-contiguous RoPE.md
|
…
|
|
|
68 - It is time to fix replace_all.md
|
…
|
|
|
69 - Allow bf16 kv-cache.md
|
…
|
|
|
70 - Fused unary_x_y.md
|
…
|
|
|
71 - iqk_mul_mat_ better srategy when nrc_y not divisible by ny.md
|
…
|
|
|
72 - iqk_mul_mat_ better iq4_nl implementation on Zen4_AVX2.md
|
…
|
|
|
73 - CUDA_ faster float -_ iq4_nl conversion.md
|
…
|
|
|
74 - IQ4_NL kv-cache on the CPU _Zen4_AVX2_ARM_NEON_.md
|
…
|
|
|
75 - Fix Q5_0 flash attention.md
|
…
|
|
|
76 - iq4_nl_ faster quantization.md
|
…
|
|
|
77 - Adding Q6_0.md
|
…
|
|
|
78 - q6_0_ Slightly faster Zen4_AVX2.md
|
…
|
|
|
79 - Do not quantize activations if not necessary.md
|
…
|
|
|
80 - Move to c_17 projectwide.md
|
…
|
|
|
81 - Cleanup scale fudge factors.md
|
…
|
|
|
83 - New SOTA quantization_ 4.25 bpw IQ4_KS.md
|
…
|
|
|
84 - Better model info.md
|
…
|
|
|
85 - IQ2_KS_ 2.1875 bpw non-linear quantization.md
|
…
|
|
|
86 - Fix and optimize iq2k Metal implementation.md
|
…
|
|
|
87 - iq3_k_ fix and optimize Metal dot product.md
|
…
|
|
|
89 - Adding IQ4_KSS_ 4.0 bpw quants.md
|
…
|
|
|
90 - iq4_ks_ faster dot product on Metal.md
|
…
|
|
|
91 - CLI - Specify GGML_TYPE to quantize for the main tensors..md
|
…
|
|
|
93 - Attempt to blindly fix Windows build failure.md
|
…
|
|
|
94 - Adding _agray3_s graph caching approach.md
|
…
|
|
|
96 - Quant strategies_ attn_q Q4 _ attn_v Q6 for Llama 3.1 Q5_K_S.md
|
…
|
|
|
97 - Bitnet_ make the scale tensors optional.md
|
…
|
|
|
98 - Avoid rebuild of GGML graph for each token.md
|
…
|
|
|
99 - Enable IQ4_NL for KV-cache in token generation using Flash Attention.md
|
…
|
|
|
101 - Enable q6_0 in flash attention.md
|
…
|
|
|
102 - Add support for Granite and GraniteMoE models.md
|
…
|
|
|
105 - Fix quantized k-cache without FA.md
|
…
|
|
|
106 - Bitnet changes.md
|
…
|
|
|
107 - Faster IQ1_BN Metal implementation.md
|
…
|
|
|
108 - Another Bitnet performance improvement on Metal.md
|
…
|
|
|
109 - Bitnet CUDA improvements.md
|
…
|
|
|
110 - Bitnet_ use the fused mul-silu in the FFN network.md
|
…
|
|
|
111 - Use fused mul - unary op also for MoE models.md
|
…
|
|
|
112 - Faster MoE inference.md
|
…
|
|
|
113 - Trellis quantization.md
|
…
|
|
|
114 - MMQ Kernel for Q6_0 _pretty please_.md
|
…
|
|
|
115 - MMQ for Q6_0.md
|
…
|
|
|
116 - Use Q6_0 instead of Q5_1 for tensors incompatible with IQ5_K_Q5_K.md
|
…
|
|
|
117 - Some minor quant strategies tweaks.md
|
…
|
|
|
118 - IQ4_NL_X4.md
|
…
|
|
|
119 - Q4_0_R4.md
|
…
|
|
|
120 - Q8_0_R4.md
|
…
|
|
|
121 - Q5_0_R4.md
|
…
|
|
|
122 - Q6_0_R4.md
|
…
|
|
|
123 - IQ4_XS_R4.md
|
…
|
|
|
124 - iq2_bn_r4_ fastest Bitnet CPU implementation on the planet.md
|
…
|
|
|
125 - R4 improvements on ARM_NEON.md
|
…
|
|
|
126 - Rename iq4_nl_x4 to iq4_nl_r4.md
|
…
|
|
|
127 - Q4_0_R4 on CUDA.md
|
…
|
|
|
128 - Faster IQ4_XS_R4 on Zen4.md
|
…
|
|
|
129 - Q4_K_R4.md
|
…
|
|
|
130 - Q6_K_R4.md
|
…
|
|
|
131 - Slightly faster Q4_K_R4 and IQ4_XS_R4 on Zen4.md
|
…
|
|
|
132 - Q5_K_R4.md
|
…
|
|
|
134 - Q3_K_R4.md
|
…
|
|
|
135 - Better ARM_NEON implementation for R4 quants.md
|
…
|
|
|
136 - Q2_K_R4.md
|
…
|
|
|
137 - Fix AVX2 implementation of iq4_nl_r4.md
|
…
|
|
|
138 - IQ4_K_R4.md
|
…
|
|
|
139 - Faster R4 quants on Zen4.md
|
…
|
|
|
141 - Q8_K_R8_ Fastest quantized matrix multiplications.md
|
…
|
|
|
142 - BF16_R16 - 16 interleaved bf16 rows.md
|
…
|
|
|
143 - Slightly faster IQ4_XS_R4 on AVX2.md
|
…
|
|
|
144 - Slightly faster IQ4_K_R4 on AVX2_Zen4.md
|
…
|
|
|
145 - IQ3_K_R4.md
|
…
|
|
|
146 - IQ2_K_R4.md
|
…
|
|
|
147 - Be able to repack tensors at run time.md
|
…
|
|
|
148 - Slightly better matrix x vector on Zen4_AVX2 for iq2_k_r4_ iq3_k_r4_ iq.md
|
…
|
|
|
149 - IQ5_K_R4.md
|
…
|
|
|
150 - IQ4_KS_R4.md
|
…
|
|
|
151 - fix typo.md
|
…
|
|
|
152 - IQ3_XXS_R4.md
|
…
|
|
|
153 - IQ3_XXS_R4.md
|
…
|
|
|
154 - IQ2_XXS_R4.md
|
…
|
|
|
155 - IQ2_XS_R4.md
|
…
|
|
|
156 - IQ2_S_R4.md
|
…
|
|
|
157 - R4 i-quants improvements.md
|
…
|
|
|
158 - Faster R4 legacy quants.md
|
…
|
|
|
161 - MSVC fixes.md
|
…
|
|
|
162 - IQ3_S_R4.md
|
…
|
|
|
163 - q4_0_r4_ Use AVX2 version for matrix x vector.md
|
…
|
|
|
168 - Falcon3 changes.md
|
…
|
|
|
169 - Be able to re-quantize MS BitNet I2_S models.md
|
…
|
|
|
170 - MoE fix for R4 quants.md
|
…
|
|
|
171 - Fix lower FA performance for even batch sizes.md
|
…
|
|
|
172 - CPU Flash Attention improvements.md
|
…
|
|
|
173 - More Flash Attention improvements.md
|
…
|
|
|
174 - On Zen4 repack fp16 models to bf16_r16.md
|
…
|
|
|
175 - Better BF16 support on AVX2.md
|
…
|
|
|
176 - Deepseek V3 support added.md
|
…
|
|
|
177 - Update chat templates.md
|
…
|
|
|
178 - Interleave 8 rows _Q8_0_ IQ4_XS_.md
|
…
|
|
|
179 - Minor performance improvements.md
|
…
|
|
|
180 - Deepseek MLA Optimizations.md
|
…
|
|
|
181 - Various.md
|
…
|
|
|
182 - Faster Q4_K_R4 and Q5_K_R4 on AVX2_Zen4.md
|
…
|
|
|
184 - Deepseek-Lite.md
|
…
|
|
|
185 - IQ1_S_R4_ better 1.5 bpw quants.md
|
…
|
|
|
186 - iq1_s_r4_ slightly faster NEON gemm_gemv.md
|
…
|
|
|
187 - IQ1_M_R4_ better 1.75 bpw quants.md
|
…
|
|
|
188 - Add optional MLA.md
|
…
|
|
|
189 - Rename q4_0_r4_ q8_0_r4 and iq4_xs_r4 to _r8.md
|
…
|
|
|
190 - cuda_ non-contiguous rms norm.md
|
…
|
|
|
191 - Add additional checks for iq1_s_r4 quantization.md
|
…
|
|
|
192 - Revert _79.md
|
…
|
|
|
193 - RPC sync.md
|
…
|
|
|
194 - Use Q8_K_128 for IQ1_S_R4 and IQ1_M_R4 matrix multiplications.md
|
…
|
|
|
195 - Deepseek MLA Optimizations V2.md
|
…
|
|
|
197 - FA_ Add option to build all FA kernels.md
|
…
|
|
|
198 - Load all MoE experts during warmup and make warmup 1 token.md
|
…
|
|
|
200 - DeepSeek FA support _CPU only_.md
|
…
|
|
|
202 - Fix imatrix overprotectiveness.md
|
…
|
|
|
204 - Fix iqk_mul_mat on AVX512 systems that are missing BF16 support.md
|
…
|
|
|
205 - Faster MLA prompt processing.md
|
…
|
|
|
206 - MLA_ allow Q8_0 K-cache for MLA.md
|
…
|
|
|
207 - Faster CPU TG for GQA models.md
|
…
|
|
|
208 - Q8_KV_ 8-bit quantization type targeting the KV cache.md
|
…
|
|
|
210 - Repack also experts.md
|
…
|
|
|
212 - Optimized GEMM_GEMV for IQ1_S.md
|
…
|
|
|
213 - Fix NEON gemm_gemv for legacy quants when row size is not divisible by .md
|
…
|
|
|
215 - Trying to fix confusion betweem HAVE_FANCY_SIMD and AVX512.md
|
…
|
|
|
216 - Hopefully this really fixes the confusion between AVX512 and FANCY_SIMD.md
|
…
|
|
|
218 - Better strategy for attention matrix multiplications when generating to.md
|
…
|
|
|
219 - Fuse MoE up and gate matrix multiplications.md
|
…
|
|
|
220 - Fix _217.md
|
…
|
|
|
225 - Examples _ Add new sweep-bench benchmark.md
|
…
|
|
|
226 - Fix compilation error with IQK_FA_ALL_QUANTS enabled.md
|
…
|
|
|
229 - Fused MoE ffn_up and ffn_gate.md
|
…
|
|
|
231 - Fix _230.md
|
…
|
|
|
232 - Give the user the option to override where model weights are stored.md
|
…
|
|
|
233 - Slightly faster CUDA MLA.md
|
…
|
|
|
234 - Faster MLA on CUDA.md
|
…
|
|
|
235 - Option to use MLA without a transposed cache.md
|
…
|
|
|
236 - Feat_lock free server.md
|
…
|
|
|
237 - Reduce size of compute buffers.md
|
…
|
|
|
238 - A better way to measure the cost of ggml_barrier.md
|
…
|
|
|
239 - SER - Smart Expert Reduction.md
|
…
|
|
|
240 - Flash MLA _CPU only_.md
|
…
|
|
|
241 - DeepSeek CUDA Flash Attention.md
|
…
|
|
|
243 - Better FlashMLA.md
|
…
|
|
|
244 - Custom quantization rules with regular expressions.md
|
…
|
|
|
246 - Faster FlashMLA prompt processing.md
|
…
|
|
|
247 - FlashMLA on CUDA.md
|
…
|
|
|
248 - Faster MoE token generation on CUDA.md
|
…
|
|
|
250 - DeepSeek imatrix stuff.md
|
…
|
|
|
251 - Try using fp32 for FlashMLA.md
|
…
|
|
|
252 - MLA-2_ Allow usage of q8_0 for KV cache on CUDA.md
|
…
|
|
|
253 - FlashMLA-2 _CPU_ faster and smaller compute buffer size.md
|
…
|
|
|
259 - Prepare wk_b tensors of DeepSeek models on the fly.md
|
…
|
|
|
260 - FlashMLA-2_ reduce compute buffer size _CUDA and CPU_.md
|
…
|
|
|
261 - Compile time option to use bf16 for quants without MMQ kernels.md
|
…
|
|
|
262 - Fix _261.md
|
…
|
|
|
264 - Make Q8_0 KV cache work with FlasMLA-2 on CUDA.md
|
…
|
|
|
265 - Allow q8_0 cache on the CPU for FlashMLA-2.md
|
…
|
|
|
268 - Prevent FlashMLA-1 from running on CUDA.md
|
…
|
|
|
269 - Fix ggml_compute_forward_dup_q.md
|
…
|
|
|
270 - Honor mmap setting when using tensor overrides.md
|
…
|
|
|
272 - Convert models to row-interleaved quants using the quantize tool.md
|
…
|
|
|
273 - FlashMLA-3_ the best of both worlds _CPU only_.md
|
…
|
|
|
274 - Specify tensor name regex for tensors to be repacked.md
|
…
|
|
|
275 - Fix bug_ missing parentheses in logical expression.md
|
…
|
|
|
276 - Add Gemma3 support _text only_.md
|
…
|
|
|
277 - Attempt to improve FlashMLA on the CPU.md
|
…
|
|
|
278 - Test transparent huge pages on Linux.md
|
…
|
|
|
279 - Fighting with cmake.md
|
…
|
|
|
280 - Native build ooption for CUDA when GGML_NATIVE is set.md
|
…
|
|
|
282 - Improve DeepSeek batched processing speed.md
|
…
|
|
|
283 - CUDA_ better MoE implementation.md
|
…
|
|
|
284 - llama-bench_ enable having different number of threads for tg and pp.md
|
…
|
|
|
287 - Is this better for DeepSeek-R1_.md
|
…
|
|
|
289 - Update sweep bench _depracating .jsonl support_.md
|
…
|
|
|
290 - mmap backed KV cache.md
|
…
|
|
|
291 - Disable Zen4 optimizations for Q8_0_Q8_0_R8.md
|
…
|
|
|
292 - Use bf16 instead of fp16 block scales for q8_1.md
|
…
|
|
|
294 - Make sure tensor row size is multiple of block size also when quantizin.md
|
…
|
|
|
295 - Quantization improvements.md
|
…
|
|
|
298 - Update gguf-py constants.md
|
…
|
|
|
299 - Additional guards for interleaved quants.md
|
…
|
|
|
301 - Fix _300.md
|
…
|
|
|
302 - Quantization improvements _2_.md
|
…
|
|
|
303 - Fix ARM_NEON build failure due to q8_2.md
|
…
|
|
|
307 - Metal_ much faster MoE prompt processing.md
|
…
|
|
|
309 - Fix GCC compilation errors on ARM.md
|
…
|
|
|
310 - Metal_ FA and FlashMLA.md
|
…
|
|
|
311 - Add -flax-vector-conversions for GCC on ARM.md
|
…
|
|
|
312 - Improved IQ2_XS quantization.md
|
…
|
|
|
313 - We need to synchronize before using device to host async memcpy.md
|
…
|
|
|
315 - Try not repacking q8_0 for FA computations.md
|
…
|
|
|
317 - Add copyright notices.md
|
…
|
|
|
318 - Use links for ggml_llama.cpp authors.md
|
…
|
|
|
320 - Guard against attempts to use MLA for non-MLA models.md
|
…
|
|
|
321 - LlaMA-4 support _text only_.md
|
…
|
|
|
324 - Correct L4 rms_norm.md
|
…
|
|
|
325 - Fix KLD precision.md
|
…
|
|
|
326 - WIP Compute per layer LIM Scores during imatrix.md
|
…
|
|
|
327 - Improved IQ1_M quantization.md
|
…
|
|
|
328 - imatrix_ collect layer influence statistics.md
|
…
|
|
|
329 - Add ability to hide imatrix details in llama-quantize.md
|
…
|
|
|
330 - Allow q8_0 KV cache for head size 256.md
|
…
|
|
|
331 - Better gemm_gemv on AVX2 fr q4_0_r8.md
|
…
|
|
|
332 - Better TG performance for GQA models _CPU_.md
|
…
|
|
|
333 - Support GLM-4-0414 models based on piDack_s mainline PR.md
|
…
|
|
|
336 - Fix termux_android build.md
|
…
|
|
|
337 - Add support for bitnet2b_2501 model.md
|
…
|
|
|
338 - BitNet adjustments.md
|
…
|
|
|
341 - Add support for Cohere2.md
|
…
|
|
|
342 - Fix LLaMA-4 attention.md
|
…
|
|
|
343 - cuda_ use switch in constexpr funcs.md
|
…
|
|
|
344 - Add GLM-4-0414 Model Support.md
|
…
|
|
|
346 - Fix FA on ARM CPUs.md
|
…
|
|
|
347 - Add ability to manually set arch flags.md
|
…
|
|
|
348 - Fix q4_1 and q5_1 on Arm.md
|
…
|
|
|
349 - Fix division by zero bug.md
|
…
|
|
|
351 - CPU FA improvements.md
|
…
|
|
|
352 - Update README.md.md
|
…
|
|
|
355 - Apply Qwen3 PR from llama.cpp.md
|
…
|
|
|
356 - Add missing enum values for qwen3 and qwen3moe.md
|
…
|
|
|
360 - Fix IQK_FA_ALL_QUANTS on AVX2.md
|
…
|
|
|
364 - Fix FA bug on AVX2.md
|
…
|
|
|
366 - Add support for new Bitnet model architecture name.md
|
…
|
|
|
368 - Trying to fix iq1_s_r4_iq1_m_r4 quantization failure.md
|
…
|
|
|
369 - cmake_ force MSVC compiler charset to utf-8.md
|
…
|
|
|
370 - CUDA_ faster FA TG for GQA models.md
|
…
|
|
|
371 - Another attempt to fix _367.md
|
…
|
|
|
374 - CUDA_ MMQ for IQ4_KS.md
|
…
|
|
|
375 - Add batch warmup to sweep-bench.md
|
…
|
|
|
377 - Support for Llama-3-Nemotron models.md
|
…
|
|
|
382 - Fix DeepSeek FA.md
|
…
|
|
|
386 - FlashMLA-3 for DeepSeek models on CUDA.md
|
…
|
|
|
390 - Fix build for Xeon Gold 6226R.md
|
…
|
|
|
391 - Fix DeepSeek q8_0 cache.md
|
…
|
|
|
392 - fix some MSVC build problem..md
|
…
|
|
|
394 - Handle incompatible DeepSeek GGUFs.md
|
…
|
|
|
400 - Fix CUDA DeepSeek FlashMLA-3 with quantized KV cache.md
|
…
|
|
|
402 - Fix missing rope_freqs with convert_hf_to_gguf.md
|
…
|
|
|
404 - TG improvements for MoE models.md
|
…
|
|
|
405 - GPU offload policy.md
|
…
|
|
|
406 - Fix race in the CUDA DeepSeek FA kernel.md
|
…
|
|
|
408 - Faster DeepSeek FA on CUDA.md
|
…
|
|
|
409 - Enable faster prompt processing with mainline llama.cpp GGUFs.md
|
…
|
|
|
410 - Better CPU FA performance for DeepSeek-Lite.md
|
…
|
|
|
411 - Fix imatrix calculation for MLA models.md
|
…
|
|
|
413 - Fix new CUDA FA on Touring.md
|
…
|
|
|
415 - Fix SER _CPU_.md
|
…
|
|
|
416 - Fix SER _CUDA_.md
|
…
|
|
|
417 - CUDA_ quantized GEMM for for IQ4_K_ IQ5_K_ IQ6_K.md
|
…
|
|
|
418 - CUDA_ quantized GEMM for for IQ2_KS_ IQ2_K_ IQ3_K.md
|
…
|
|
|
421 - Fix standard attention on the CPU.md
|
…
|
|
|
422 - Adding IQ5_KS - 5.25 bpw quants.md
|
…
|
|
|
424 - Adding forgotten template instance for iq5_ks.md
|
…
|
|
|
426 - IQ5_KS_R4_ row-interleaved IQ5_KS.md
|
…
|
|
|
427 - Fix AVX2 implementation of IQ4_K_ IQ4_KS_ IQ5_K_ IQ6_K.md
|
…
|
|
|
428 - Zen4_ Faster PP for IQ2_KS_ IQ4_KS_ IQ5_KS.md
|
…
|
|
|
429 - Option to enable or disable the CPU FA kernels.md
|
…
|
|
|
430 - Disable multi-add for now.md
|
…
|
|
|
431 - Forgotten MMQ ref and typo.md
|
…
|
|
|
435 - Refactor iqk_mul_mat.cpp.md
|
…
|
|
|
438 - Another attempt to fix the illegal memory access bug.md
|
…
|
|
|
439 - Bug fixes from mainline.md
|
…
|
|
|
441 - Trellis quants with CPU inference.md
|
…
|
|
|
442 - CUDA call tracer.md
|
…
|
|
|
443 - Streamline a bit the quant strategies.md
|
…
|
|
|
444 - gguf-split _ update.md
|
…
|
|
|
445 - Fix typo in non-AVX2 code branch.md
|
…
|
|
|
446 - Fix bug in MMVQ kernel.md
|
…
|
|
|
448 - Fix MSVC compilation.md
|
…
|
|
|
449 - Legacy quants conversion schemes in convert_hf_to_gguf.py.md
|
…
|
|
|
453 - Faster IQ3_KT and IQ4_KT.md
|
…
|
|
|
454 - Add support for FP8 GGUF creation and re-quantization _WIP_.md
|
…
|
|
|
457 - Remove GGML_IQK_MUL_MAT option.md
|
…
|
|
|
458 - Add missing gguf-py constants.md
|
…
|
|
|
460 - aarch64 kernels for KT quants.md
|
…
|
|
|
461 - CUDA implementation for IQ2_K_R4_ IQ3_K_R4_ IQ4_K_R4_ IQ5_K_R4.md
|
…
|
|
|
462 - CUDA GEMM and GEMV for IQ4_KS_R4 and IQ5_KS_R4.md
|
…
|
|
|
465 - Set cache_prompt default to true.md
|
…
|
|
|
468 - Minor _2_ iq2_ks TG performance improvement on CUDA.md
|
…
|
|
|
469 - Replace MLA-specific KV cache with the standard KV cache.md
|
…
|
|
|
470 - Send _DONE_ for OAI compatibility.md
|
…
|
|
|
471 - NEON implementation for trellis quants.md
|
…
|
|
|
473 - Replace MLA-specific KV cache with the standard KV cache V2.md
|
…
|
|
|
475 - Metal implementatio for the trellis quants..md
|
…
|
|
|
478 - forgotten refs and typo.md
|
…
|
|
|
480 - Rpc improvement.md
|
…
|
|
|
481 - Webui improvement.md
|
…
|
|
|
482 - Trellis quants_ faster CPU prompt processing.md
|
…
|
|
|
483 - convert_hf_to_gguf.py _ conversion from hf weights to Q6_0.md
|
…
|
|
|
484 - BF16 Trellis implementation.md
|
…
|
|
|
486 - Adding the XTC sampler.md
|
…
|
|
|
487 - Make sure MMVQ is supported before using it.md
|
…
|
|
|
488 - Faster CPU prompt processing for Trellis quants and MoE models.md
|
…
|
|
|
489 - Adding top-n-sigma sampler.md
|
…
|
|
|
492 - CUDA implementation for IQ1_S_R4.md
|
…
|
|
|
493 - MMQ implementation for IQ4_KS_R4 and IQ5_KS_R4.md
|
…
|
|
|
494 - IQ1_M_R4 CUDA implementation.md
|
…
|
|
|
495 - Check if ffn_up and ffn_gate are of the same type before using fmoe.md
|
…
|
|
|
496 - Quick hack_ add the MLA flag to llama_hparams.md
|
…
|
|
|
497 - Make prompt cache saving and restoring MLA aware.md
|
…
|
|
|
501 - Fix _499.md
|
…
|
|
|
502 - Add an endpoint that lists all the saved prompt caches to server.md
|
…
|
|
|
504 - Add DRY and fix the server to use other new samplers..md
|
…
|
|
|
505 - New IQ4_KT trellis implementation.md
|
…
|
|
|
506 - Fix non rpc build error.md
|
…
|
|
|
508 - Fix Compile error _C2668_.md
|
…
|
|
|
509 - Docs update.md
|
…
|
|
|
510 - Update News section of readme.md
|
…
|
|
|
511 - New IQ2_KT.md
|
…
|
|
|
512 - Add top n sigma sampler in webui and other webui fix.md
|
…
|
|
|
513 - add dry sampler.md
|
…
|
|
|
515 - IQ2_XXS_ much faster CPU prompt processing.md
|
…
|
|
|
516 - Much faster iq3_xxs GEMM via repacking to q8_0_r8 _AVX2_.md
|
…
|
|
|
517 - IQ1_S_ much faster CPU prompt processing.md
|
…
|
|
|
518 - IQ3_S_ much faster CPU prompt processing.md
|
…
|
|
|
520 - Better strategy for GPU offload.md
|
…
|
|
|
524 - Perhaps a slightly better GEMV version for IQ2_XXS_ IQ3_XXS_ IQ3_S.md
|
…
|
|
|
525 - Faster CPU prompt processing for Q4_K and Q5_K.md
|
…
|
|
|
528 - Fix bug introduced in _524_525.md
|
…
|
|
|
529 - New IQ2_KT_ IQ3_KT and IQ4_KT_ V2.md
|
…
|
|
|
531 - Much faster CPU prompt processing _part 1_.md
|
…
|
|
|
533 - Much faster CPU prompt processing _part 2_.md
|
…
|
|
|
534 - Much faster CPU prompt processing _part 3_.md
|
…
|
|
|
535 - Minor readme update.md
|
…
|
|
|
536 - Fix KT Neon _ ARM typo.md
|
…
|
|
|
537 - Update CMakeLists.txt to fix NDEBUG handling.md
|
…
|
|
|
540 - Fix missed block_q8_x2 bf16 -_ i16 change.md
|
…
|
|
|
541 - Perhaps slightly faster trellis quants.md
|
…
|
|
|
542 - Fix NEON build.md
|
…
|
|
|
544 - New integer trellis on ARM_NEON.md
|
…
|
|
|
546 - Faster ARM_NEON GEMM implementation for legacy quants.md
|
…
|
|
|
547 - build_ add script to simplify build_test workflow for Android.md
|
…
|
|
|
549 - Much faster prompt processing for IQK quants _ARM_NEON_.md
|
…
|
|
|
550 - Much faster prompt processing for I-quants _ARM_NEON_.md
|
…
|
|
|
552 - Much faster prompt processing for k-quants _ARM_NEON_.md
|
…
|
|
|
553 - Much faster prompt processing for IQ1_S and IQ1_M on ARM_NEON.md
|
…
|
|
|
554 - Update README.md to add quickstart section.md
|
…
|
|
|
555 - Add Falcon-Edge support.md
|
…
|
|
|
557 - CUDA_ MMQ for iqX_r4 quants.md
|
…
|
|
|
558 - Add mikupad to ik_llama as an alternative WebUI.md
|
…
|
|
|
559 - Use cuBLAS for large batches and quants with block size 16.md
|
…
|
|
|
560 - Remove what appears to be unnecessary asserts in ggml_cuda_cpy.md
|
…
|
|
|
563 - Merge vulkan code from mainline up to commit of 6_28_2025.md
|
…
|
|
|
565 - add hunyuan moe support for 561.md
|
…
|
|
|
566 - Adding IQ3_KS quants.md
|
…
|
|
|
567 - Minor CUDA PP speed improvement.md
|
…
|
|
|
569 - Conditionally disable fused ops when building with Vulkan enabled.md
|
…
|
|
|
570 - Remove duplicate_misplaced cmake find_package for Vulkan.md
|
…
|
|
|
571 - Fix CMakeLists.md
|
…
|
|
|
573 - Support for dots.llm1 models.md
|
…
|
|
|
574 - Change KQ mask padding to 64.md
|
…
|
|
|
577 - Vulkan_ fused rms norm.md
|
…
|
|
|
578 - Do not crash when there is no DRY sampler.md
|
…
|
|
|
579 - Fix debug build failure with RPC off.md
|
…
|
|
|
580 - Vulkan_ add GGML_OP_FUSED_MUL_UNARY.md
|
…
|
|
|
581 - Vulkan_ Disable multi-add for now.md
|
…
|
|
|
582 - Vulkan_ adding GGML_OP_MULTI_ADD implementation.md
|
…
|
|
|
583 - Adding forgotten file.md
|
…
|
|
|
584 - Vulkan_ flash attention for DeepSeek models.md
|
…
|
|
|
585 - Special handling of Seed Coder FIM tokens.md
|
…
|
|
|
587 - Fix crash when there is no DRY sampler.md
|
…
|
|
|
588 - Fix server crash when there is no DRY sampler.md
|
…
|
|
|
589 - CUDA_ small PP performance improvement for MoE models.md
|
…
|
|
|
592 - Another minor readme update.md
|
…
|
|
|
593 - Faster prompt processing for IQ2_KS_ IQ2_K_ IQ2_K_R4.md
|
…
|
|
|
595 - CUDA_ Faster prompt processing for several quantization types.md
|
…
|
|
|
598 - Vulkan_ iquants and flash attention split_k_reduce improvement.md
|
…
|
|
|
602 - Adding IQ2_KL.md
|
…
|
|
|
603 - Check if MMQ should be used before using it.md
|
…
|
|
|
604 - Fix attn_v conditionality when quantizing..md
|
…
|
|
|
606 - Add iq3_ks to constants.py.md
|
…
|
|
|
607 - vulkan_ support softmax_FA batch and broadcast.md
|
…
|
|
|
608 - Vulkan_ a fresh start.md
|
…
|
|
|
609 - Added kimi-k2 support _ported from llama.cpp_.md
|
…
|
|
|
610 - q8_k_r8_ experimental AVX512 version.md
|
…
|
|
|
611 - Bump GGML_MAX_CONTEXTS to allow loading more shards.md
|
…
|
|
|
612 - kimi-k2 convert script and chat template.md
|
…
|
|
|
616 - Adding IQ1_KT - 1.75 bpw SOTA quants.md
|
…
|
|
|
617 - Fixup kimi-k2 convert indentation.md
|
…
|
|
|
618 - Webui_ New Features for Conversations_ Settings_ and Chat Messages.md
|
…
|
|
|
620 - Bump Windows max open files from 512 to 2048.md
|
…
|
|
|
622 - Add GGML_MAX_CONTEXTS definition in CMakeLists.txt.md
|
…
|
|
|
624 - Quantization tweaks.md
|
…
|
|
|
628 - _Draft_ Function calling support for Kimi-K2.md
|
…
|
|
|
630 - GEMM for IQ1_M.md
|
…
|
|
