Code

sync : llama.cpp#1089

Code

Merged

ggerganov merged 38 commits intomasterggml-org/ggml:masterfrom sync-llama.cpp-25-01-29ggml-org/ggml:sync-llama.cpp-25-01-29Copy head branch name to clipboardJan 29, 2025

+3,641-1,475

Commits

Commits on Jan 29, 2025

• SYCL: Add gated linear attention kernel (llama/11175)

  Show description for 28ac740
  

  qnixsynapse

  authored and

  ggerganov

  committed
  28ac740

  Copy full SHA for 28ac740

• RoPE: fix back, CUDA support for back + noncont. (llama/11240)

  Show description for 904a095
  

  JohannesGaessler

  authored and

  ggerganov

  committed
  904a095

  Copy full SHA for 904a095

• fix: ggml: fix vulkan-shaders-gen build (llama/10448)

  Show description for ba41673
  

  sparkleholic

  authored and

  ggerganov

  committed
  ba41673

  Copy full SHA for ba41673

• vulkan: scale caching for k quants + misc fixes (llama/11081)

  Show description for 0cab8e0
  

  netrunnereve

  authored and

  ggerganov

  committed
  0cab8e0

  Copy full SHA for 0cab8e0

• ggml: aarch64: implement SVE kernels for q4_K_q8_K vector dot (llama/11227)

  Show description for e8ed706
  

  fj-y-saito

  and

  ggerganov

  committed
  e8ed706

  Copy full SHA for e8ed706

• CUDA: backwards pass for misc. ops, add tests (llama/11257)

  Show description for 9b5d224
  

  JohannesGaessler

  authored and

  ggerganov

  committed
  9b5d224

  Copy full SHA for 9b5d224

• vulkan: optimize coopmat2 q2_k dequant function (llama/11130)

  

  jeffbolznv

  authored and

  ggerganov

  committed
  934f7ec

  Copy full SHA for 934f7ec

• vulkan: optimize coopmat2 q4_k/q5_k dequant functions. (llama/11206)

  Show description for be82ddf
  

  jeffbolznv

  authored and

  ggerganov

  committed
  be82ddf

  Copy full SHA for be82ddf

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

  Show description for ba0e978
  

  jeffbolznv

  authored and

  ggerganov

  committed
  ba0e978

  Copy full SHA for ba0e978

• rpc : early register backend devices (llama/11262)

  Show description for 72a5ae0
  

  rgerganov

  authored and

  ggerganov

  committed
  72a5ae0

  Copy full SHA for 72a5ae0

• vulkan: fix coopmat2 flash attention for non-contiguous inputs (llama/11281)

  Show description for 782151b
  

  jeffbolznv

  authored and

  ggerganov

  committed
  782151b

  Copy full SHA for 782151b

• cmake : add sanitizer flags for llama.cpp (llama/11279)

  Show description for dc90506
  

  ggerganov

  and

  JohannesGaessler

  committed
  dc90506

  Copy full SHA for dc90506

• SYCL: Introducing memory host pool (llama/11251)

  Show description for fbaffa1
  

  s-Nick

  authored and

  ggerganov

  committed
  fbaffa1

  Copy full SHA for fbaffa1

• vulkan: fix coopmat2 validation failures (llama/11284)

  Show description for 3cf35e1
  

  jeffbolznv

  authored and

  ggerganov

  committed
  3cf35e1

  Copy full SHA for 3cf35e1

• metal : fix out-of-bounds write (llama/11314)

  Show description for 4ec5cf7
  ggerganov

  committed
  4ec5cf7

  Copy full SHA for 4ec5cf7

• rpc : better caching of the base buffer pointer (llama/11331)

  Show description for 54c1e2f
  

  rgerganov

  authored and

  ggerganov

  committed
  54c1e2f

  Copy full SHA for 54c1e2f

• vulkan: fix diag_mask_inf (llama/11323)

  Show description for d578c62
  

  jeffbolznv

  authored and

  ggerganov

  committed
  d578c62

  Copy full SHA for d578c62

• vulkan: sort shaders for more deterministic binary (llama/11315)

  Show description for 4bec73f
  

  jeffbolznv

  authored and

  ggerganov

  committed
  4bec73f

  Copy full SHA for 4bec73f

• Vulkan-run-test: fix mmq_wg_denoms (llama/11343)

  Show description for 05a0c1d
  

  AMD-dwang

  authored and

  ggerganov

  committed
  05a0c1d

  Copy full SHA for 05a0c1d

• tests: fix some mul_mat test gaps (llama/11375)

  Show description for 027cb54
  

  jeffbolznv

  authored and

  ggerganov

  committed
  027cb54

  Copy full SHA for 027cb54

• cmake : avoid -march=native when reproducible build is wanted (llama/11366)

  Show description for 5eb4acd
  

  bmwiedemann

  authored and

  ggerganov

  committed
  5eb4acd

  Copy full SHA for 5eb4acd

• CPU/CUDA: fix (GQA) mul mat back, add CUDA support (llama/11380)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  cd3261d

  Copy full SHA for cd3261d

• rocBLAS: Avoid fp32->fp16->fp32 conversion on cdna (llama/11356)

  

  IMbackK

  authored and

  ggerganov

  committed
  8c59363

  Copy full SHA for 8c59363

• CUDA: fix FP16 cuBLAS GEMM (llama/11396)

  

  JohannesGaessler

  authored and

  ggerganov

  committed
  7957144

  Copy full SHA for 7957144

• hip : Add hipGraph and VMM support to ROCM (llama/11362)

  Show description for c8e6c31
  

  IMbackK

  authored and

  ggerganov

  committed
  c8e6c31

  Copy full SHA for c8e6c31

• Hip: disable VMM on hip as it seams that it dosent work in some configurations (llama/11420)

  

  IMbackK

  authored and

  ggerganov

  committed
  2a0f12c

  Copy full SHA for 2a0f12c

• vulkan: compile shaders on-demand (llama/11406)

  Show description for d886b3b
  

  jeffbolznv

  authored and

  ggerganov

  committed
  d886b3b

  Copy full SHA for d886b3b

• cmake: add ggml find package (llama/11369)

  Show description for 0e85d87
  

  bandoti

  and

  ggerganov

  committed
  0e85d87

  Copy full SHA for 0e85d87

• metal : use residency sets (llama/11427)

  Show description for a756ee5
  ggerganov

  committed
  a756ee5

  Copy full SHA for a756ee5

• metal: Handle null returned from MTLCreateSystemDefaultDevice() (llama/11441)

  Show description for bc64584
  

  booxter

  authored and

  ggerganov

  committed
  bc64584

  Copy full SHA for bc64584

• AMD: parse the architecture as supplied by gcnArchName (llama/11244)

  Show description for ec4d3e7
  

  Haus1

  authored and

  ggerganov

  committed
  ec4d3e7

  Copy full SHA for ec4d3e7

• SYCL : SOFTMAX F16 mask support and other fixes (llama/11261)

  Show description for 2aa6088
  

  qnixsynapse

  authored and

  ggerganov

  committed
  2aa6088

  Copy full SHA for 2aa6088

• cmake : don't fail on GGML_CPU=OFF (llama/11457)

  

  someone13574

  authored and

  ggerganov

  committed
  8ac9155

  Copy full SHA for 8ac9155

• HIP: Only call rocblas_initialize on rocblas versions with the multiple instantation bug (llama/11080)

  Show description for 807b5f2
  

  sARY77

  authored and

  ggerganov

  committed
  807b5f2

  Copy full SHA for 807b5f2

• HIP: Supress transformation warning in softmax.cu

  Show description for df283f7
  

  IMbackK

  authored and

  ggerganov

  committed
  df283f7

  Copy full SHA for df283f7

• sync : llama.cpp

  Show description for 5bcbe65
  ggerganov

  committed
  5bcbe65

  Copy full SHA for 5bcbe65

• scripts : sync cmake

  ggerganov

  committed
  5ce4ce2

  Copy full SHA for 5ce4ce2

• cmake : sync new file

  Show description for 41ae935
  ggerganov

  committed
  41ae935

  Copy full SHA for 41ae935