github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #8058] CUSTOM_CPU_FLAGS="" / non avx2 build #30912

New Issue
Closed
opened 2026-04-22 10:54:11 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-22 10:54:11 -05:00
Owner
Copy Link

Originally created by @agreppin on GitHub (Dec 12, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/8058

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

sequel of #7640 Error: POST predict: Post "http://127.0.0.1:42623/completion": EOF

here is the command I used to build ollama:

make -j16 HIP_ARCHS=gfx1102 CUSTOM_CPU_FLAGS="" RUNNER_TARGETS=rocm dist

here is the relevant log part:

...
Dec 12 00:45:57 prox1 ollama[1632910]: ggml_cuda_init: found 1 ROCm devices:
Dec 12 00:45:57 prox1 ollama[1632910]:   Device 0: AMD Radeon™ RX 7600 XT, compute capability 11.0, VMM: no
Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: ggml ctx size =    0.43 MiB
Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: offloading 32 repeating layers to GPU
Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: offloading non-repeating layers to GPU
Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: offloaded 33/33 layers to GPU
Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors:      ROCm0 buffer size =  1903.13 MiB
Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors:        CPU buffer size =    98.44 MiB
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: n_ctx      = 8192
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: n_batch    = 2048
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: n_ubatch   = 512
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: flash_attn = 0
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: freq_base  = 10000000.0
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: freq_scale = 1
Dec 12 00:45:58 prox1 ollama[1632910]: llama_kv_cache_init:      ROCm0 KV buffer size =  2560.00 MiB
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: KV self size  = 2560.00 MiB, K (f16): 1280.00 MiB, V (f16): 1280.00 MiB
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model:  ROCm_Host  output buffer size =     0.79 MiB
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model:      ROCm0 compute buffer size =   548.00 MiB
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model:  ROCm_Host compute buffer size =    21.01 MiB
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: graph nodes  = 1254
Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: graph splits = 2
Dec 12 01:32:55 prox1 ollama[1642790]: time=2024-12-12T01:32:55.520+01:00 level=INFO source=server.go:594 msg="llama runner started in 3.26 seconds"
Dec 12 01:32:55 prox1 ollama[1642790]: [GIN] 2024/12/12 - 01:32:55 | 200 |  3.426053264s |       127.0.0.1 | POST     "/api/generate"
Dec 12 01:33:06 prox1 ollama[1642790]: SIGILL: illegal instruction
Dec 12 01:33:06 prox1 ollama[1642790]: PC=0x7513232be656 m=5 sigcode=2
Dec 12 01:33:06 prox1 ollama[1642790]: signal arrived during cgo execution
Dec 12 01:33:06 prox1 ollama[1642790]: instruction bytes: 0xc4 0xe2 0x79 0x13 0xc0 0xc5 0xf0 0x57 0xc9 0xc5 0xf8 0x2e 0xc1 0xf 0x9b 0xc0
...

objdump -d ./dist/linux-amd64/lib/ollama/runners/rocm/ollama_llama_server shows avx instructions:

   d5d1c:	c4 e2 7d 78 c8       	vpbroadcastb %xmm0,%ymm1
   d646d:	c4 e2 7d 78 c8       	vpbroadcastb %xmm0,%ymm1
  27005a:	c4 e3 45 06 c6 03    	vperm2f128 $0x3,%ymm6,%ymm7,%ymm0
  27006a:	c4 e3 55 06 cc 03    	vperm2f128 $0x3,%ymm4,%ymm5,%ymm1
  270108:	c4 e3 7d 06 e0 00    	vperm2f128 $0x0,%ymm0,%ymm0,%ymm4
  270116:	c4 e3 45 06 d7 11    	vperm2f128 $0x11,%ymm7,%ymm7,%ymm2
  270242:	c4 e3 5d 02 e2 f0    	vpblendd $0xf0,%ymm2,%ymm4,%ymm4
...

Idem for ollama and libggml_rocm.so, .

I will try to find the origin of the byte-codes sequence '0xc4 0xe2 0x79 0x13 0xc0 0xc5 0xf0 0x57 0xc9 0xc5 0xf8 0x2e 0xc1 0xf 0x9b 0xc0'. If someone know how to do this, let us know.

Open questions:

  • origin of those avx instructions, ollama code or ROCm libs ?
  • how can I start the runner in a debugger ?

OS

Linux

GPU

AMD

CPU

Other

Ollama version

0.5.2-rc1-0-g6a6328a, 0.5.0-rc1-6-g1841309

Originally created by @agreppin on GitHub (Dec 12, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/8058 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? sequel of #7640 Error: POST predict: Post "http://127.0.0.1:42623/completion": EOF here is the command I used to build ollama: ```sh make -j16 HIP_ARCHS=gfx1102 CUSTOM_CPU_FLAGS="" RUNNER_TARGETS=rocm dist ``` here is the relevant log part: ``` ... Dec 12 00:45:57 prox1 ollama[1632910]: ggml_cuda_init: found 1 ROCm devices: Dec 12 00:45:57 prox1 ollama[1632910]: Device 0: AMD Radeon™ RX 7600 XT, compute capability 11.0, VMM: no Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: ggml ctx size = 0.43 MiB Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: offloading 32 repeating layers to GPU Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: offloading non-repeating layers to GPU Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: offloaded 33/33 layers to GPU Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: ROCm0 buffer size = 1903.13 MiB Dec 12 00:45:57 prox1 ollama[1632910]: llm_load_tensors: CPU buffer size = 98.44 MiB Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: n_ctx = 8192 Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: n_batch = 2048 Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: n_ubatch = 512 Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: flash_attn = 0 Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: freq_base = 10000000.0 Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: freq_scale = 1 Dec 12 00:45:58 prox1 ollama[1632910]: llama_kv_cache_init: ROCm0 KV buffer size = 2560.00 MiB Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: KV self size = 2560.00 MiB, K (f16): 1280.00 MiB, V (f16): 1280.00 MiB Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: ROCm_Host output buffer size = 0.79 MiB Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: ROCm0 compute buffer size = 548.00 MiB Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: ROCm_Host compute buffer size = 21.01 MiB Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: graph nodes = 1254 Dec 12 00:45:58 prox1 ollama[1632910]: llama_new_context_with_model: graph splits = 2 Dec 12 01:32:55 prox1 ollama[1642790]: time=2024-12-12T01:32:55.520+01:00 level=INFO source=server.go:594 msg="llama runner started in 3.26 seconds" Dec 12 01:32:55 prox1 ollama[1642790]: [GIN] 2024/12/12 - 01:32:55 | 200 | 3.426053264s | 127.0.0.1 | POST "/api/generate" Dec 12 01:33:06 prox1 ollama[1642790]: SIGILL: illegal instruction Dec 12 01:33:06 prox1 ollama[1642790]: PC=0x7513232be656 m=5 sigcode=2 Dec 12 01:33:06 prox1 ollama[1642790]: signal arrived during cgo execution Dec 12 01:33:06 prox1 ollama[1642790]: instruction bytes: 0xc4 0xe2 0x79 0x13 0xc0 0xc5 0xf0 0x57 0xc9 0xc5 0xf8 0x2e 0xc1 0xf 0x9b 0xc0 ... ``` objdump -d `./dist/linux-amd64/lib/ollama/runners/rocm/ollama_llama_server` shows avx instructions: ``` d5d1c: c4 e2 7d 78 c8 vpbroadcastb %xmm0,%ymm1 d646d: c4 e2 7d 78 c8 vpbroadcastb %xmm0,%ymm1 27005a: c4 e3 45 06 c6 03 vperm2f128 $0x3,%ymm6,%ymm7,%ymm0 27006a: c4 e3 55 06 cc 03 vperm2f128 $0x3,%ymm4,%ymm5,%ymm1 270108: c4 e3 7d 06 e0 00 vperm2f128 $0x0,%ymm0,%ymm0,%ymm4 270116: c4 e3 45 06 d7 11 vperm2f128 $0x11,%ymm7,%ymm7,%ymm2 270242: c4 e3 5d 02 e2 f0 vpblendd $0xf0,%ymm2,%ymm4,%ymm4 ... ``` Idem for ollama and libggml_rocm.so, . I will try to find the origin of the byte-codes sequence '0xc4 0xe2 0x79 0x13 0xc0 0xc5 0xf0 0x57 0xc9 0xc5 0xf8 0x2e 0xc1 0xf 0x9b 0xc0'. If someone know how to do this, let us know. Open questions: - origin of those avx instructions, ollama code or ROCm libs ? - how can I start the runner in a debugger ? ### OS Linux ### GPU AMD ### CPU Other ### Ollama version 0.5.2-rc1-0-g6a6328a, 0.5.0-rc1-6-g1841309
GiteaMirror added the buildbug labels 2026-04-22 10:54:11 -05:00
GiteaMirror commented 2026-04-22 10:54:13 -05:00
Author
Owner
Copy Link

@agreppin commented on GitHub (Dec 12, 2024):

Here is a full log of that problem, including strace:

ollama-strace-0.5.2-rc1-0-g6a6328a.log.gz

The illegal instruction is after a call to ioctl:

...
1652949 [0000741920524ded] ioctl(3, AMDKFD_IOC_ALLOC_MEMORY_OF_GPU, 0x7418a3df5f30) = 0
1652949 [0000741920524ded] ioctl(3, AMDKFD_IOC_MAP_MEMORY_TO_GPU, 0x7418a3df5f70) = 0
1652949 [000074191e4d65d6] --- SIGILL {si_signo=SIGILL, si_code=ILL_ILLOPN, si_addr=0x74191e4d65d6} ---
...

Seems to me that the amdgpu driver is the problem. (not sure, see edit 1)

Here is the script used to produce the log file:

#!/bin/sh
set -eux
LOG=ollama-strace.log
exec >$LOG
exec 2>&1
printf '# %s: ft=strace\n' vi
PWD=~/src/ollama
export LD_LIBRARY_PATH=$PWD/dist/linux-amd64/lib/ollama/runners/rocm:$PWD/dist/linux-amd64-rocm/lib/ollama
ldd $PWD/dist/linux-amd64/lib/ollama/runners/rocm/ollama_llama_server
env -i HOME=$HOME strace -f -i -v -s 16384 -o $LOG -A $PWD/dist/linux-amd64/bin/ollama serve 2>&1

edit 1:
librocblas.so.4.2.60204 contains a F16C (CPU feature) instruction, likely the cause of SIGILL:

c4 e2 79 13 c0       	vcvtph2ps %xmm0,%xmm0

found the issue for rocm-6.2.4, appears to be fixed in rocm-6.3.0:
https://github.com/ROCm/rocBLAS/issues/1422#issuecomment-2089026056

edit 2:
librocblas.so.4.2.60204 rebuilt without -mf16c, the issue seems fixed for my case !

but that question remains open:

  • ollama, libggml_rocm.so & rocm/ollama_llama_server have avx instructions, why ?

one is in libggml_rocm.so / llama/ggml-cuda/ggml-cuda.cu / ggml_cuda_init()

c4 e3 75 18 c9 01       vinsertf128 $0x1,%xmm1,%ymm1,%ymm1

edit 3:

make -f make/Makefile.rocm HIP_ARCHS=gfx1102 CUSTOM_CPU_FLAGS="" RUNNER_TARGETS=rocm print-GPU_COMPILER_CFLAGS print-GPU_COMPILER_CXXFLAGS print-GPU_COMPILER_CUFLAGS print-GPU_RUNNER_ARCH_FLAGS print-GPU_RUNNER_CPU_FLAGS

GPU_COMPILER_CFLAGS= -fPIC -D_GNU_SOURCE
GPU_COMPILER_CXXFLAGS= -fPIC -D_GNU_SOURCE
GPU_COMPILER_CUFLAGS=-fPIC -Wno-unused-function -std=gnu++17  -mf16c -mfma -c -O3 -DGGML_USE_CUDA -DGGML_BUILD=1 -DGGML_BACKEND_BUILD=1 -DGGML_SHARED=1 -DGGML_BACKEND_SHARED=1 -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_MMV_Y=1 -DGGML_SCHED_MAX_COPIES=4 -DGGML_USE_HIP -DGGML_USE_LLAMAFILE -DHIP_FAST_MATH -D__HIP_PLATFORM_AMD__=1 -D__HIP_ROCclr__=1 -DNDEBUG -DK_QUANTS_PER_ITERATION=2 -D_CRT_SECURE_NO_WARNINGS -D_GNU_SOURCE -D_XOPEN_SOURCE=600 -DUSE_PROF_API=1 -std=gnu++14 -x hip -mllvm=-amdgpu-early-inline-all=true -mllvm=-amdgpu-function-calls=false -Wno-expansion-to-defined -Wno-invalid-noreturn -Wno-ignored-attributes -Wno-pass-failed -Wno-deprecated-declarations -Wno-unused-result -I./llama/
GPU_RUNNER_ARCH_FLAGS=--offload-arch=gfx1102
GPU_RUNNER_CPU_FLAGS=

so, -mf16c needs to be removed from GPU_COMPILER_CUFLAGS for non avx builds, CUSTOM_CPU_FLAGS="".
This is strange that my Xeon CPU E5-2680 has avx, but no f16c.

edit 4:
-mfma needs to be removed too from make/Makefile.rocm.

<!-- gh-comment-id:2538742241 --> @agreppin commented on GitHub (Dec 12, 2024): Here is a full log of that problem, including strace: [ollama-strace-0.5.2-rc1-0-g6a6328a.log.gz](https://github.com/user-attachments/files/18111136/ollama-strace-0.5.2-rc1-0-g6a6328a.log.gz) The illegal instruction is after a call to ioctl: ``` ... 1652949 [0000741920524ded] ioctl(3, AMDKFD_IOC_ALLOC_MEMORY_OF_GPU, 0x7418a3df5f30) = 0 1652949 [0000741920524ded] ioctl(3, AMDKFD_IOC_MAP_MEMORY_TO_GPU, 0x7418a3df5f70) = 0 1652949 [000074191e4d65d6] --- SIGILL {si_signo=SIGILL, si_code=ILL_ILLOPN, si_addr=0x74191e4d65d6} --- ... ``` Seems to me that the amdgpu driver is the problem. (not sure, see edit 1) Here is the script used to produce the log file: ```sh #!/bin/sh set -eux LOG=ollama-strace.log exec >$LOG exec 2>&1 printf '# %s: ft=strace\n' vi PWD=~/src/ollama export LD_LIBRARY_PATH=$PWD/dist/linux-amd64/lib/ollama/runners/rocm:$PWD/dist/linux-amd64-rocm/lib/ollama ldd $PWD/dist/linux-amd64/lib/ollama/runners/rocm/ollama_llama_server env -i HOME=$HOME strace -f -i -v -s 16384 -o $LOG -A $PWD/dist/linux-amd64/bin/ollama serve 2>&1 ``` edit 1: librocblas.so.4.2.60204 contains a F16C (CPU feature) instruction, likely the cause of SIGILL: ``` c4 e2 79 13 c0 vcvtph2ps %xmm0,%xmm0 ``` found the issue for rocm-6.2.4, appears to be fixed in rocm-6.3.0: https://github.com/ROCm/rocBLAS/issues/1422#issuecomment-2089026056 edit 2: librocblas.so.4.2.60204 rebuilt without -mf16c, the issue seems fixed for my case ! but that question remains open: - ollama, libggml_rocm.so & rocm/ollama_llama_server have avx instructions, why ? one is in libggml_rocm.so / llama/ggml-cuda/ggml-cuda.cu / ggml_cuda_init() ``` c4 e3 75 18 c9 01 vinsertf128 $0x1,%xmm1,%ymm1,%ymm1 ``` edit 3: ```sh make -f make/Makefile.rocm HIP_ARCHS=gfx1102 CUSTOM_CPU_FLAGS="" RUNNER_TARGETS=rocm print-GPU_COMPILER_CFLAGS print-GPU_COMPILER_CXXFLAGS print-GPU_COMPILER_CUFLAGS print-GPU_RUNNER_ARCH_FLAGS print-GPU_RUNNER_CPU_FLAGS GPU_COMPILER_CFLAGS= -fPIC -D_GNU_SOURCE GPU_COMPILER_CXXFLAGS= -fPIC -D_GNU_SOURCE GPU_COMPILER_CUFLAGS=-fPIC -Wno-unused-function -std=gnu++17 -mf16c -mfma -c -O3 -DGGML_USE_CUDA -DGGML_BUILD=1 -DGGML_BACKEND_BUILD=1 -DGGML_SHARED=1 -DGGML_BACKEND_SHARED=1 -DGGML_CUDA_DMMV_X=32 -DGGML_CUDA_MMV_Y=1 -DGGML_SCHED_MAX_COPIES=4 -DGGML_USE_HIP -DGGML_USE_LLAMAFILE -DHIP_FAST_MATH -D__HIP_PLATFORM_AMD__=1 -D__HIP_ROCclr__=1 -DNDEBUG -DK_QUANTS_PER_ITERATION=2 -D_CRT_SECURE_NO_WARNINGS -D_GNU_SOURCE -D_XOPEN_SOURCE=600 -DUSE_PROF_API=1 -std=gnu++14 -x hip -mllvm=-amdgpu-early-inline-all=true -mllvm=-amdgpu-function-calls=false -Wno-expansion-to-defined -Wno-invalid-noreturn -Wno-ignored-attributes -Wno-pass-failed -Wno-deprecated-declarations -Wno-unused-result -I./llama/ GPU_RUNNER_ARCH_FLAGS=--offload-arch=gfx1102 GPU_RUNNER_CPU_FLAGS= ``` so, **-mf16c** needs to be removed from GPU_COMPILER_CUFLAGS for non avx builds, CUSTOM_CPU_FLAGS="". This is strange that my Xeon CPU E5-2680 has avx, but no [f16c](https://en.wikipedia.org/wiki/F16C). edit 4: **-mfma** needs to be removed too from make/Makefile.rocm.
GiteaMirror commented 2026-04-22 10:54:13 -05:00
Author
Owner
Copy Link

@agreppin commented on GitHub (Dec 15, 2024):

Successful non avx build result with the following mods:

  • removed -mf16c -mfma from make/Makefile.rocm
  • recompiled ROCm 6.2.4 with some mods (the hard part)
  • using CUSTOM_CPU_FLAGS="tune=generic" (go build tag ignored, Makefile.rocm => -mtune=generic)

The last traces of avx instructions are in:

bin/ollama:
<countbody>:
<indexbytebody>:
<math.archExp.abi0>:
<crypto/sha512.blockAVX2.abi0>:
<crypto/sha256.block.abi0>:
<crypto/sha1.blockAVX2.abi0>:
<vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Open.abi0>:
<vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Seal.abi0>:
<github.com/apache/arrow/go/arrow/memory._memset_avx2.abi0>:

runners/rocm/ollama_llama_server:
<countbody>:
<indexbytebody>:
<crypto/sha512.blockAVX2.abi0>:
<crypto/sha256.block.abi0>:
<crypto/sha1.blockAVX2.abi0>:
<vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Open.abi0>:
<vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Seal.abi0>:

... wich seems golang code paths for the CPU the executable will run on. So, the last question is answered.

Notes: the ROCm online doc says:

ROCm requires CPUs that support PCIe™ atomics. Modern CPUs after the release of 1st generation AMD Zen CPU and Intel™ Haswell support PCIe atomics.

gcc -march=znver1 & -march=haswell both have AVX2, F16C & FMA. You can verify with:

gcc -march=haswell -Q --help=target -xc /dev/null
<!-- gh-comment-id:2543953768 --> @agreppin commented on GitHub (Dec 15, 2024): Successful non avx build result with the following mods: - removed -mf16c -mfma from make/Makefile.rocm - recompiled ROCm 6.2.4 with some mods (the hard part) - using CUSTOM_CPU_FLAGS="tune=generic" (go build tag ignored, Makefile.rocm => -mtune=generic) The last traces of avx instructions are in: ``` bin/ollama: <countbody>: <indexbytebody>: <math.archExp.abi0>: <crypto/sha512.blockAVX2.abi0>: <crypto/sha256.block.abi0>: <crypto/sha1.blockAVX2.abi0>: <vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Open.abi0>: <vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Seal.abi0>: <github.com/apache/arrow/go/arrow/memory._memset_avx2.abi0>: runners/rocm/ollama_llama_server: <countbody>: <indexbytebody>: <crypto/sha512.blockAVX2.abi0>: <crypto/sha256.block.abi0>: <crypto/sha1.blockAVX2.abi0>: <vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Open.abi0>: <vendor/golang.org/x/crypto/chacha20poly1305.chacha20Poly1305Seal.abi0>: ``` ... wich seems golang code paths for the CPU the executable will run on. So, the last question is answered. Notes: the [ROCm online doc](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/reference/system-requirements.html#cpu-support) says: > ROCm requires CPUs that support PCIe™ atomics. Modern CPUs after the release of 1st generation AMD Zen CPU and Intel™ Haswell support PCIe atomics. gcc -march=znver1 & -march=haswell both have AVX2, F16C & FMA. You can verify with: ```sh gcc -march=haswell -Q --help=target -xc /dev/null ```
GiteaMirror commented 2026-04-22 10:54:13 -05:00
Author
Owner
Copy Link

@dhiltgen commented on GitHub (Apr 9, 2025):

We've revamped the build which now builds multiple CPU optimized libraries distinct from the GPU libraries.

https://github.com/ollama/ollama/blob/main/docs/development.md

<!-- gh-comment-id:2790758875 --> @dhiltgen commented on GitHub (Apr 9, 2025): We've revamped the build which now builds multiple CPU optimized libraries distinct from the GPU libraries. https://github.com/ollama/ollama/blob/main/docs/development.md
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label bug build
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#30912
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?