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

[PR #15811] docs: add RDNA4 / gfx1201 ROCm build instructions to development.md #77610

New Issue
Open
opened 2026-05-05 10:16:49 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-05-05 10:16:49 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/15811
Author: @LukeLamb
Created: 4/25/2026
Status: 🔄 Open

Base: main ← Head: docs-rdna4-gfx1201

📝 Commits (1)

  • e7cb21f docs: add RDNA4 / gfx1201 ROCm build instructions

📊 Changes

1 file changed (+35 additions, -0 deletions)

View changed files

📝 docs/development.md (+35 -0)

📄 Description

Summary

The official ollama-linux-amd64 binary in 0.20.6 ships CPU and CUDA backends only — /usr/local/lib/ollama/ has cuda_v12/, cuda_v13/, and CPU variants, but no rocm/ runner directory. RDNA4 users (RX 9000-series, Radeon AI PRO R9700, gfx1200/gfx1201) hitting the install script today get a CPU-only experience and have no in-tree pointer telling them they need to build from source.

The source already supports gfx1200/gfx1201 (CMakeLists.txt L141 regex and the "ROCm 7" preset in CMakePresets.json), so this is a docs-only change — no code touched. It adds a short subsection to docs/development.md covering:

  • The reason a from-source build is needed today (linking back to #10676 for when the official Linux ROCm build resumes)
  • ROCm version requirements (≥ 6.3 for gfx1201; ≥ 7.0 recommended for hipBLASLt-tuned kernels)
  • The exact cmake --preset 'ROCm 7' -DAMDGPU_TARGETS=gfx1201 … command set
  • A verification step (grep "inference compute" in serve output, looking for library=ROCm compute=gfx1201)

Verified end-to-end

Tested on Ubuntu 24.04 + kernel 6.17 + ROCm 7.2.1 against an AMD Radeon AI PRO R9700:

  • library=ROCm compute=gfx1201 name=ROCm0 description="AMD Radeon Graphics" pci_id=0000:0d:00.0 type=discrete total="31.9 GiB" available="31.8 GiB"
  • 33/33 layers offloaded
  • 92.99 tok/s on llama3.1:8b-q4_K_M (vs 88.59 tok/s on Vulkan/RADV and 8.77 tok/s on CPU)

Test plan

  • Cloned ollama/ollama HEAD, ran the documented build commands verbatim, confirmed library=ROCm compute=gfx1201 in the journal
  • Confirmed link targets resolve: #10676 (build pipeline tracker), #14927 (matching-symptom bug + repro), CMakePresets.json "ROCm 7" preset
  • Markdown lint: section sits between the existing "## Linux" body and "## MLX Engine (Optional)" heading; doesn't disturb adjacent content

Refs

  • #14927 — RDNA4 gfx1201 not detected, falls back to CPU (the matching-symptom bug)
  • #10430 — canonical RDNA4 9070/9070 XT support tracker
  • #10676 — draft PR to bump official build pipeline to ROCm 6.4+ (stalled on Windows)
  • #13236 — R9700-specific ROCm discovery timeout

🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.

## 📋 Pull Request Information **Original PR:** https://github.com/ollama/ollama/pull/15811 **Author:** [@LukeLamb](https://github.com/LukeLamb) **Created:** 4/25/2026 **Status:** 🔄 Open **Base:** `main` ← **Head:** `docs-rdna4-gfx1201` --- ### 📝 Commits (1) - [`e7cb21f`](https://github.com/ollama/ollama/commit/e7cb21fdb8572538cb059e1af797773d544dfff8) docs: add RDNA4 / gfx1201 ROCm build instructions ### 📊 Changes **1 file changed** (+35 additions, -0 deletions) <details> <summary>View changed files</summary> 📝 `docs/development.md` (+35 -0) </details> ### 📄 Description ## Summary The official `ollama-linux-amd64` binary in 0.20.6 ships CPU and CUDA backends only — `/usr/local/lib/ollama/` has `cuda_v12/`, `cuda_v13/`, and CPU variants, but no `rocm/` runner directory. RDNA4 users (RX 9000-series, Radeon AI PRO R9700, gfx1200/gfx1201) hitting the install script today get a CPU-only experience and have no in-tree pointer telling them they need to build from source. The source already supports gfx1200/gfx1201 (`CMakeLists.txt` L141 regex and the `"ROCm 7"` preset in `CMakePresets.json`), so this is a docs-only change — no code touched. It adds a short subsection to `docs/development.md` covering: - The reason a from-source build is needed today (linking back to #10676 for when the official Linux ROCm build resumes) - ROCm version requirements (≥ 6.3 for gfx1201; ≥ 7.0 recommended for hipBLASLt-tuned kernels) - The exact `cmake --preset 'ROCm 7' -DAMDGPU_TARGETS=gfx1201 …` command set - A verification step (`grep "inference compute"` in serve output, looking for `library=ROCm compute=gfx1201`) ## Verified end-to-end Tested on Ubuntu 24.04 + kernel 6.17 + ROCm 7.2.1 against an AMD Radeon AI PRO R9700: - `library=ROCm compute=gfx1201 name=ROCm0 description="AMD Radeon Graphics" pci_id=0000:0d:00.0 type=discrete total="31.9 GiB" available="31.8 GiB"` - 33/33 layers offloaded - 92.99 tok/s on `llama3.1:8b-q4_K_M` (vs 88.59 tok/s on Vulkan/RADV and 8.77 tok/s on CPU) ## Test plan - [x] Cloned `ollama/ollama` HEAD, ran the documented build commands verbatim, confirmed `library=ROCm compute=gfx1201` in the journal - [x] Confirmed link targets resolve: #10676 (build pipeline tracker), #14927 (matching-symptom bug + repro), `CMakePresets.json` "ROCm 7" preset - [x] Markdown lint: section sits between the existing "## Linux" body and "## MLX Engine (Optional)" heading; doesn't disturb adjacent content ## Refs - #14927 — RDNA4 gfx1201 not detected, falls back to CPU (the matching-symptom bug) - #10430 — canonical RDNA4 9070/9070 XT support tracker - #10676 — draft PR to bump official build pipeline to ROCm 6.4+ (stalled on Windows) - #13236 — R9700-specific ROCm discovery timeout --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>
GiteaMirror added the pull-request label 2026-05-05 10:16:50 -05:00
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 pull-request
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#77610
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?