github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #8305] Speed ten times slower than llamafile #67374

New Issue
Closed
opened 2026-05-04 10:08:09 -05:00 by GiteaMirror · 7 comments
GiteaMirror commented 2026-05-04 10:08:09 -05:00
Owner
Copy Link

Originally created by @ErfolgreichCharismatisch on GitHub (Jan 4, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/8305

Originally assigned to: @dhiltgen on GitHub.

What is the issue?

Llamafile is much faster on cpu than ollama, what takes ollama 33 minutes takes llamafile 3 minutes with the same model.
llamafile crashes unfortunately after reusing it and spins its wheels staying at 100% CPU for hours.

I'd rather use a stable ollama, but you must work on speed on CPU

OS

Linux

CPU

Intel

Originally created by @ErfolgreichCharismatisch on GitHub (Jan 4, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/8305 Originally assigned to: @dhiltgen on GitHub. ### What is the issue? Llamafile is much faster on cpu than ollama, what takes ollama 33 minutes takes llamafile 3 minutes with the same model. llamafile crashes unfortunately after reusing it and spins its wheels staying at 100% CPU for hours. I'd rather use a stable ollama, but you must work on speed on CPU ### OS Linux ### CPU Intel
GiteaMirror added the bugperformanceneeds more info labels 2026-05-04 10:08:10 -05:00
GiteaMirror commented 2026-05-04 10:08:13 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 4, 2025):

How did you compare them?

<!-- gh-comment-id:2571277470 --> @rick-github commented on GitHub (Jan 4, 2025): How did you compare them?
GiteaMirror commented 2026-05-04 10:08:14 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 4, 2025):

services:
  ollama:
    image: ollama/ollama:${OLLAMA_DOCKER_TAG-latest}
    volumes:
      - ${OLLAMA_MODELS-../ollama-data}:/root/.ollama
    ports:
      - 11434:11434

  llamafile:
    image: iverly/llamafile-docker:${LLAMAFILE_DOCKER_TAG-latest}
    volumes:
      - ${LLAMAFILE_MODEL-./llama3.1-8b.gguf}:/model
    ports:
      - 8081:8080

#!/usr/bin/env python3

import requests
import time

prompt = "why is the sky blue"
base_url = "http://localhost:{port}/v1/chat/completions"
model = "llama3.1:8b"

ports = {
  "11434": "ollama-0.5.4",
  "8081": "llamafile-0.8.17",
}

data = {
  "model": model,
  "messages": [{"role":"user","content":prompt}],
  "temperature":0,
  "seed":-1,
}

for port in ports.keys():
  tokens = 0
  elapsed = 0
  for _ in range(10):
    start = time.time()
    response = requests.post(base_url.format(port=port), json=data).json()
    end = time.time()
    tokens = tokens + response['usage']['completion_tokens']
    elapsed = elapsed + end - start
  print(f"{ports[port]:<16}: {tokens} tokens in {elapsed:.2f} seconds, {tokens/elapsed:.2f} tokens/s")

$ docker compose up -d
[+] Running 2/3
 ⠏ Network 8305_default        Created                                                                                                                          0.9s 
 ✔ Container 8305-llamafile-1  Started                                                                                                                          0.8s 
 ✔ Container 8305-ollama-1     Started                                                                                                                          0.8s 
$ ./8305.py 
ollama-0.5.4    : 4535 tokens in 363.67 seconds, 12.47 tokens/s
llamafile-0.8.17: 4250 tokens in 334.54 seconds, 12.70 tokens/s

$ docker compose logs | grep threads
llamafile-1  | {"function":"server_cli","level":"INFO","line":2918,"msg":"system info","n_threads":8,"n_threads_batch":8,"system_info":"AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | ","tid":"12066432","timestamp":1736004445,"total_threads":24}
ollama-1     | time=2025-01-04T15:27:47.600Z level=INFO source=server.go:376 msg="starting llama server" cmd="/usr/lib/ollama/runners/cpu_avx2/ollama_llama_server runner --model /root/.ollama/models/blobs/sha256-8eeb52dfb3bb9aefdf9d1ef24b3bdbcfbe82238798c4b918278320b6fcef18fe --ctx-size 8192 --batch-size 512 --threads 8 --no-mmap --parallel 4 --port 42999"
ollama-1     | time=2025-01-04T15:27:47.604Z level=INFO source=runner.go:946 msg=system info="CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | cgo(gcc)" threads=8
<!-- gh-comment-id:2571341863 --> @rick-github commented on GitHub (Jan 4, 2025): ```yaml services: ollama: image: ollama/ollama:${OLLAMA_DOCKER_TAG-latest} volumes: - ${OLLAMA_MODELS-../ollama-data}:/root/.ollama ports: - 11434:11434 llamafile: image: iverly/llamafile-docker:${LLAMAFILE_DOCKER_TAG-latest} volumes: - ${LLAMAFILE_MODEL-./llama3.1-8b.gguf}:/model ports: - 8081:8080 ``` ```python #!/usr/bin/env python3 import requests import time prompt = "why is the sky blue" base_url = "http://localhost:{port}/v1/chat/completions" model = "llama3.1:8b" ports = { "11434": "ollama-0.5.4", "8081": "llamafile-0.8.17", } data = { "model": model, "messages": [{"role":"user","content":prompt}], "temperature":0, "seed":-1, } for port in ports.keys(): tokens = 0 elapsed = 0 for _ in range(10): start = time.time() response = requests.post(base_url.format(port=port), json=data).json() end = time.time() tokens = tokens + response['usage']['completion_tokens'] elapsed = elapsed + end - start print(f"{ports[port]:<16}: {tokens} tokens in {elapsed:.2f} seconds, {tokens/elapsed:.2f} tokens/s") ``` ```console $ docker compose up -d [+] Running 2/3 ⠏ Network 8305_default Created 0.9s ✔ Container 8305-llamafile-1 Started 0.8s ✔ Container 8305-ollama-1 Started 0.8s $ ./8305.py ollama-0.5.4 : 4535 tokens in 363.67 seconds, 12.47 tokens/s llamafile-0.8.17: 4250 tokens in 334.54 seconds, 12.70 tokens/s ``` ```console $ docker compose logs | grep threads llamafile-1 | {"function":"server_cli","level":"INFO","line":2918,"msg":"system info","n_threads":8,"n_threads_batch":8,"system_info":"AVX = 1 | AVX_VNNI = 1 | AVX2 = 1 | AVX512 = 0 | AVX512_VBMI = 0 | AVX512_VNNI = 0 | AVX512_BF16 = 0 | FMA = 1 | NEON = 0 | SVE = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | SSSE3 = 1 | VSX = 0 | MATMUL_INT8 = 0 | LLAMAFILE = 1 | ","tid":"12066432","timestamp":1736004445,"total_threads":24} ollama-1 | time=2025-01-04T15:27:47.600Z level=INFO source=server.go:376 msg="starting llama server" cmd="/usr/lib/ollama/runners/cpu_avx2/ollama_llama_server runner --model /root/.ollama/models/blobs/sha256-8eeb52dfb3bb9aefdf9d1ef24b3bdbcfbe82238798c4b918278320b6fcef18fe --ctx-size 8192 --batch-size 512 --threads 8 --no-mmap --parallel 4 --port 42999" ollama-1 | time=2025-01-04T15:27:47.604Z level=INFO source=runner.go:946 msg=system info="CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | LLAMAFILE = 1 | AARCH64_REPACK = 1 | cgo(gcc)" threads=8 ```
GiteaMirror commented 2026-05-04 10:08:16 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 5, 2025):

From https://justine.lol/matmul/, llamafile can improve prompt processing on CPU but not completion processing. Since llamafile uses llama.cpp as the backend, as the llamafile improvements are upstreamed to llama.cpp (https://github.com/ggerganov/llama.cpp/pull/6414, https://github.com/ggerganov/llama.cpp/pull/6412), ollama benefits automatically. Ollama is already receiving this as shown in the LLAMAFILE setting in the system info.

<!-- gh-comment-id:2571452369 --> @rick-github commented on GitHub (Jan 5, 2025): From https://justine.lol/matmul/, llamafile can improve prompt processing on CPU but not completion processing. Since llamafile uses llama.cpp as the backend, as the llamafile improvements are upstreamed to llama.cpp (https://github.com/ggerganov/llama.cpp/pull/6414, https://github.com/ggerganov/llama.cpp/pull/6412), ollama benefits automatically. Ollama is already receiving this as shown in the LLAMAFILE setting in the system info.
GiteaMirror commented 2026-05-04 10:08:18 -05:00
Author
Owner
Copy Link

@jart commented on GitHub (Jan 5, 2025):

Even if ollama is using my matmul code, it's not going to help unless you tune your build system to target all the various microarchitectures at runtime. See for example:

This is what lets llamafile not only have really fast matrix multiplication, but also dispatch to the most appropriate compilation for the specific microarchitecture of your cpu.

<!-- gh-comment-id:2571536150 --> @jart commented on GitHub (Jan 5, 2025): Even if ollama is using my matmul code, it's not going to help unless you tune your build system to target all the various microarchitectures at runtime. See for example: - https://github.com/Mozilla-Ocho/llamafile/blob/c2933599286e6d58e63916e4494fdd3b30363ce7/llamafile/BUILD.mk#L70-L129 - https://github.com/Mozilla-Ocho/llamafile/blob/c2933599286e6d58e63916e4494fdd3b30363ce7/llamafile/sgemm.cpp#L26-L131 This is what lets llamafile not only have really fast matrix multiplication, but also dispatch to the most appropriate compilation for the specific microarchitecture of your cpu.
GiteaMirror commented 2026-05-04 10:08:18 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 5, 2025):

Thanks, good to know.

<!-- gh-comment-id:2571544339 --> @rick-github commented on GitHub (Jan 5, 2025): Thanks, good to know.
GiteaMirror commented 2026-05-04 10:08:19 -05:00
Author
Owner
Copy Link

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

@ErfolgreichCharismatisch do you still see the performance slow-down on recent versions? We now generate multiple CPU architectures for x86

<!-- gh-comment-id:2790844800 --> @dhiltgen commented on GitHub (Apr 9, 2025): @ErfolgreichCharismatisch do you still see the performance slow-down on recent versions? We now generate multiple CPU architectures for x86
GiteaMirror commented 2026-05-04 10:08:20 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Oct 3, 2025):

Going to close this as stale. We can reopen if need be.

<!-- gh-comment-id:3367281249 --> @pdevine commented on GitHub (Oct 3, 2025): Going to close this as stale. We can reopen if need be.
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 needs more info performance
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#67374
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?