Georgi Gerganov fd1234cb46 llama : add gpt-oss (#15091) ...

* oai moe

* compat with new checkpoint

* add attn sink impl

* add rope scaling yarn

* logits match with latest transformers code

* wip chat template

* rm trailing space

* use ggml_scale_bias

* rm redundant is_swa_all

* convert interleaved gate_up

* graph : fix activation function to match reference (#7)

* vocab : handle o200k_harmony special tokens

* ggml : add attention sinks support (#1)

* llama : add attn sinks

* ggml : add attn sinks

* cuda : add attn sinks

* vulkan : add support for sinks in softmax

remove unnecessary return

* ggml : add fused swiglu_oai op (#11)

* ggml : add fused swiglu_oai op

* Update ggml/src/ggml-cpu/ops.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

* update CUDA impl

* cont : metal impl

* add vulkan impl

* test-backend-ops : more test cases, clean up

* llama : remove unfused impl

* remove extra lines

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: slaren <slarengh@gmail.com>

* repack mxfp4 upon conversion

* clean up a bit

* enable thinking

* add quick hack to render only some special tokens

* fix bf16 conversion

* remove vocab hack

* webui ok

* support chat parsing for gpt-oss

* fix webui

* direct mapping mxfp4, FINALLY

* force using mxfp4

* properly use lazy tensor

* ggml : add mxfp4

ggml : use e8m0 conversion instead of powf

Co-authored-by: Diego Devesa <slarengh@gmail.com>

change kvalues_mxfp4 table to match e2m1 (#6)

metal : remove quantization for now (not used)

cuda : fix disabled CUDA graphs due to ffn moe bias

vulkan : add support for mxfp4

cont : add cm2 dequant

* ggml : add ggml_add_id (#13)

* ggml : add ggml_add_id

* add cuda impl

* llama : add weight support check for add_id

* perf opt

* add vulkan impl

* rename cuda files

* add metal impl

* allow in-place ggml_add_id

* llama : keep biases on CPU with --cpu-moe

* llama : fix compile error

ggml-ci

* cuda : add fallback for __nv_cvt_e8m0_to_bf16raw

ggml-ci

* cleanup

ggml-ci

* sycl : fix supports_op for MXFP4

ggml-ci

* fix Unknown reasoning format

* ggml-cpu : fix AVX build

ggml-ci

* fix hip build

ggml-ci

* cuda : add mxfp4 dequantization support for cuBLAS

ggml-ci

* ggml-cpu : fix mxfp4 fallback definitions for some architectures

ggml-ci

* cuda : fix version required for __nv_cvt_e8m0_to_bf16raw

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
Co-authored-by: slaren <slarengh@gmail.com>

2025-08-05 22:10:36 +03:00

..

acc.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

add_id.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

add.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

argmax.comp

vulkan: implement several ops relevant for ggml_opt (#11769)

2025-02-17 07:55:57 +01:00

argsort.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

clamp.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

CMakeLists.txt

vulkan: Fix GGML_VULKAN_SHADER_DEBUG_INFO (#14427)

2025-06-27 22:35:30 -05:00

concat.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

contig_copy.comp

vulkan: Add bfloat16 support (#12554)

2025-05-01 20:49:39 +02:00

conv2d_dw.comp

sync : ggml (#13268)

2025-05-02 20:54:30 +03:00

conv2d_mm.comp

vulkan: Use coopmat2 for conv2d (#14982)

2025-08-03 14:23:57 +02:00

conv_transpose_1d.comp

ggml-vulkan: adds support for op CONV_TRANSPOSE_1D (#13813)

2025-06-04 22:02:00 +02:00

copy_from_quant.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

copy_to_quant.comp

vulkan: add RTE variants for glu/add/sub/mul/div (#14653)

2025-07-15 21:32:11 +02:00

copy.comp

vulkan: Add bfloat16 support (#12554)

2025-05-01 20:49:39 +02:00

cos.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

count_equal.comp

vulkan: implement several ops relevant for ggml_opt (#11769)

2025-02-17 07:55:57 +01:00

dequant_f32.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

dequant_funcs_cm2.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

dequant_funcs.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

dequant_head.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

dequant_iq1_m.comp

vulkan: fix warnings (#13626)

2025-05-20 21:35:16 +00:00

dequant_iq1_s.comp

vulkan: initial support for IQ1_S and IQ1_M quantizations (#11528)

2025-02-15 09:01:40 +01:00

dequant_iq2_s.comp

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

2025-01-29 18:29:39 +01:00

dequant_iq2_xs.comp

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

2025-01-29 18:29:39 +01:00

dequant_iq2_xxs.comp

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

2025-01-29 18:29:39 +01:00

dequant_iq3_s.comp

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

2025-01-29 18:29:39 +01:00

dequant_iq3_xxs.comp

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

2025-01-29 18:29:39 +01:00

dequant_iq4_nl.comp

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

2025-01-29 18:29:39 +01:00

dequant_iq4_xs.comp

vulkan: initial support for IQ4_XS quantization (#11501)

2025-02-06 07:09:59 +01:00

dequant_mxfp4.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

dequant_q2_k.comp

vulkan: fix noncontig check for mat_mul_id splitting (#14683)

2025-07-15 21:51:09 +02:00

dequant_q3_k.comp

vulkan: fix noncontig check for mat_mul_id splitting (#14683)

2025-07-15 21:51:09 +02:00

dequant_q4_0.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

dequant_q4_1.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

dequant_q4_k.comp

vulkan: fix noncontig check for mat_mul_id splitting (#14683)

2025-07-15 21:51:09 +02:00

dequant_q5_0.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

dequant_q5_1.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

dequant_q5_k.comp

vulkan: fix noncontig check for mat_mul_id splitting (#14683)

2025-07-15 21:51:09 +02:00

dequant_q6_k.comp

vulkan: fix noncontig check for mat_mul_id splitting (#14683)

2025-07-15 21:51:09 +02:00

dequant_q8_0.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

diag_mask_inf.comp

vulkan: fix diag_mask_inf (#11323)

2025-01-23 08:01:17 +01:00

div.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

flash_attn_base.comp

vulkan: Handle updated FA dim2/3 definition (#14518)

2025-07-05 09:26:04 +02:00

flash_attn_cm1.comp

vulkan: Handle updated FA dim2/3 definition (#14518)

2025-07-05 09:26:04 +02:00

flash_attn_cm2.comp

vulkan: Handle updated FA dim2/3 definition (#14518)

2025-07-05 09:26:04 +02:00

flash_attn_split_k_reduce.comp

vulkan: optimize flash attention split_k_reduce (#14554)

2025-07-08 20:11:42 +02:00

flash_attn.comp

vulkan: Handle updated FA dim2/3 definition (#14518)

2025-07-05 09:26:04 +02:00

geglu_erf.comp

ggml : implement GEGLU_ERF and GEGLU_QUICK ops (#14445)

2025-07-03 23:07:22 +02:00

geglu_quick.comp

ggml : implement GEGLU_ERF and GEGLU_QUICK ops (#14445)

2025-07-03 23:07:22 +02:00

geglu.comp

ggml : implement REGLU/GEGLU/SWIGLU ops (#14158)

2025-06-29 11:04:10 +02:00

gelu_erf.comp

add GELU_ERF (#14455)

2025-07-01 10:14:21 +02:00

gelu_quick.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

gelu.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

generic_binary_head.comp

vulkan: add RTE variants for glu/add/sub/mul/div (#14653)

2025-07-15 21:32:11 +02:00

generic_head.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

generic_unary_head.comp

vulkan: support copy from f32 to q4_0/q4_1/q5_0/q5_1/q8_0/iq4_nl (#11166)

2025-01-16 22:47:10 +01:00

get_rows_quant.comp

vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (#11595)

2025-02-28 09:42:52 +01:00

get_rows.comp

vulkan: Add bfloat16 support (#12554)

2025-05-01 20:49:39 +02:00

glu_head.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

glu_main.comp

ggml : implement REGLU/GEGLU/SWIGLU ops (#14158)

2025-06-29 11:04:10 +02:00

group_norm.comp

vulkan: fix group_norm (#10496)

2024-11-26 16:45:05 +01:00

im2col.comp

vulkan/cuda: Fix im2col when KW!=KH (#14789)

2025-07-21 13:35:40 +02:00

l2_norm.comp

llama: Add support for RWKV v7 architecture (#12412)

2025-03-18 07:27:50 +08:00

leaky_relu.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

mul_mat_split_k_reduce.comp

vulkan: optimize and reenable split_k (#10637)

2024-12-03 20:29:54 +01:00

mul_mat_vec_base.comp

vulkan: optimize mul_mat for small values of N (#10991)

2024-12-30 18:27:11 +01:00

mul_mat_vec_iq1_m.comp

vulkan: initial support for IQ1_S and IQ1_M quantizations (#11528)

2025-02-15 09:01:40 +01:00

mul_mat_vec_iq1_s.comp

vulkan: initial support for IQ1_S and IQ1_M quantizations (#11528)

2025-02-15 09:01:40 +01:00

mul_mat_vec_iq2_s.comp

vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472)

2025-03-21 20:27:47 +01:00

mul_mat_vec_iq2_xs.comp

vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (#11595)

2025-02-28 09:42:52 +01:00

mul_mat_vec_iq2_xxs.comp

vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (#11595)

2025-02-28 09:42:52 +01:00

mul_mat_vec_iq3_s.comp

vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472)

2025-03-21 20:27:47 +01:00

mul_mat_vec_iq3_xxs.comp

vulkan: add specific MMV kernels for IQ2 and IQ3 quants + optimizations (#11595)

2025-02-28 09:42:52 +01:00

mul_mat_vec_nc.comp

vulkan: Support ne[3]>1 in noncontig matrix-vector multiply (#15015)

2025-08-02 10:48:30 +02:00

mul_mat_vec_p021.comp

vulkan: Optimize mul_mat_vec p021 and nc shaders (#12505)

2025-03-22 09:40:11 +01:00

mul_mat_vec_q2_k.comp

mat vec double buffer (#12188)

2025-03-10 19:28:11 +00:00

mul_mat_vec_q3_k.comp

mat vec double buffer (#12188)

2025-03-10 19:28:11 +00:00

mul_mat_vec_q4_k.comp

vulkan: scale caching for k quants + misc fixes (#11081)

2025-01-15 19:50:13 +00:00

mul_mat_vec_q5_k.comp

vulkan: scale caching for k quants + misc fixes (#11081)

2025-01-15 19:50:13 +00:00

mul_mat_vec_q6_k.comp

mat vec double buffer (#12188)

2025-03-10 19:28:11 +00:00

mul_mat_vec.comp

vulkan: Add bfloat16 support (#12554)

2025-05-01 20:49:39 +02:00

mul_mm_cm2.comp

vulkan: optimizations for deepseek prompt processing (#14555)

2025-07-12 11:51:58 +02:00

mul_mm.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

mul_mmq_funcs.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

mul_mmq.comp

vulkan: Allow up to 4096 elements for mul_mat_id row_ids (#13326)

2025-05-09 09:23:41 +02:00

mul.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

norm.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

opt_step_adamw.comp

vulkan: implement several ops relevant for ggml_opt (#11769)

2025-02-17 07:55:57 +01:00

pad.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

pool2d.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

quantize_q8_1.comp

Vulkan: Add DP4A MMQ and Q8_1 quantization shader (#12135)

2025-03-31 14:37:01 +02:00

reglu.comp

ggml : implement REGLU/GEGLU/SWIGLU ops (#14158)

2025-06-29 11:04:10 +02:00

relu.comp

vulkan: Additional type support for unary, binary, and copy (#13266)

2025-05-04 07:17:16 +02:00

repeat_back.comp

vulkan: implement several ops relevant for ggml_opt (#11769)

2025-02-17 07:55:57 +01:00

repeat.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

rms_norm_back.comp

vulkan: implement more backpropagation operators (#11914)

2025-02-25 12:04:45 +01:00

rms_norm.comp

vulkan: fix rms_norm_mul to handle broadcasting dim0 (#14817)

2025-07-22 17:35:21 +02:00

roll.comp

vulkan : implement ggml_roll (ggml/1290)

2025-07-12 14:25:44 +03:00

rope_head.comp

vulkan: add RTE variants for glu/add/sub/mul/div (#14653)

2025-07-15 21:32:11 +02:00

rope_multi.comp

vulkan : fix rope with partial rotation and non-cont src (#14582)

2025-07-08 15:21:21 +02:00

rope_neox.comp

vulkan : fix rope with partial rotation and non-cont src (#14582)

2025-07-08 15:21:21 +02:00

rope_norm.comp

vulkan : fix rope with partial rotation and non-cont src (#14582)

2025-07-08 15:21:21 +02:00

rope_vision.comp

vulkan: support multi/vision rope, and noncontiguous rope (#11902)

2025-02-16 08:52:23 +01:00

rte.comp

vulkan: add RTE variants for glu/add/sub/mul/div (#14653)

2025-07-15 21:32:11 +02:00

scale.comp

ggml : add ggml_scale_bias (#14417)

2025-07-09 18:16:12 +02:00

sigmoid.comp

vulkan: Additional type support for unary, binary, and copy (#13266)

2025-05-04 07:17:16 +02:00

silu_back.comp

vulkan: implement more backpropagation operators (#11914)

2025-02-25 12:04:45 +01:00

silu.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

sin.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

soft_max_back.comp

vulkan: implement more backpropagation operators (#11914)

2025-02-25 12:04:45 +01:00

soft_max.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

square.comp

vulkan: Use push constant offset to handle misaligned descriptors (#10987)

2024-12-29 09:35:11 +01:00

sub.comp

vulkan: implement several ops relevant for ggml_opt (#11769)

2025-02-17 07:55:57 +01:00

sum_rows.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

swiglu_oai.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

swiglu.comp

ggml : implement REGLU/GEGLU/SWIGLU ops (#14158)

2025-06-29 11:04:10 +02:00

tanh.comp

vulkan: Additional type support for unary, binary, and copy (#13266)

2025-05-04 07:17:16 +02:00

test_bfloat16_support.comp

vulkan: Add bfloat16 support (#12554)

2025-05-01 20:49:39 +02:00

test_coopmat2_support.comp

vulkan: compile a test shader in cmake to check for coopmat2 support (#10713)

2024-12-08 09:05:55 +01:00

test_coopmat_support.comp

Disable GL_KHR_cooperative_matrix Vulkan extension if not available. (#11117)

2025-01-08 09:18:13 +01:00

test_integer_dot_support.comp

Vulkan: Add DP4A MMQ and Q8_1 quantization shader (#12135)

2025-03-31 14:37:01 +02:00

timestep_embedding.comp

ggml : build backends as libraries (#10256)

2024-11-14 18:04:35 +01:00

types.comp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

upscale.comp

vulkan : remove unused vars (#0)

2025-07-12 14:25:44 +03:00

vulkan-shaders-gen.cpp

llama : add gpt-oss (#15091)

2025-08-05 22:10:36 +03:00

wkv6.comp

rwkv6: add wkv6 support for Vulkan backend (#10829)

2024-12-16 22:00:46 +01:00

wkv7.comp

llama: Add support for RWKV v7 architecture (#12412)

2025-03-18 07:27:50 +08:00