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 #14021] [CLOSED] Add native support for AMD RDNA 4 (gfx1201 / Radeon AI Pro R9700) #14473

New Issue
Closed
opened 2026-04-13 00:55:12 -05:00 by GiteaMirror · 0 comments
GiteaMirror commented 2026-04-13 00:55:12 -05:00
Owner
Copy Link

📋 Pull Request Information

Original PR: https://github.com/ollama/ollama/pull/14021
Author: @yalbik
Created: 2/2/2026
Status: ❌ Closed

Base: main ← Head: ROCm-7

📝 Commits (1)

  • 136b3bb Add RDNA4/ROCm-7 support. Tested on Radeon AI Pro R9700 in Windows.

📊 Changes

4 files changed (+99 additions, -10 deletions)

View changed files

📝 discover/runner.go (+42 -1)
📝 docs/gpu.mdx (+16 -6)
📝 envconfig/config.go (+2 -0)
📝 llm/server.go (+39 -3)

📄 Description

** REQUIRES ROCm 7.1 on build agents **

This PR adds native discovery and build support for the RDNA 4 architecture (gfx1201). Currently, RDNA 4 cards fail discovery due to a timeout in the HIP handshake and a lack of architecture whitelisting. It was developed with the assistance of Claude Opus 4.5.

Changes:

  • Added gfx1201 to the supported AMD GPU list in gpu/amd_common.go.
  • Updated CMakePresets.json to include gfx1201 in AMDGPU_TARGETS.
  • Resolved a library pathing issue where TensileLibrary_lazy_gfx1201.dat was not being correctly located in the rocblas subdirectory.

Testing:
Verified on Windows 11 with a Radeon AI Pro R9700 and ROCm 7.1.1. Model offloading is now successful.

Primary Issues Addressed:
ollama/ollama#13236: ROCm GPU discovery times out on AMD Radeon AI PRO R9700 (gfx1201).

Why: Discovery logic hangs for 30 seconds and then falls back to 0 layers/CPU. This fix resolves this by whitelisting the architecture.

ollama/ollama#12908: ROCM Library not found, wrong location for gfx1201.

Why: This addresses the pathing error where the system looks in hipblaslt instead of rocblas for the RDNA 4 .dat files.

ollama/ollama#13908: RDNA4 support.

Why: This is a broader, more recent community request for native RDNA 4 support on Windows and Linux without using the HSA_OVERRIDE hack.

Related/Indirect Issues:
ollama/ollama#13085: Is AMD's AI Pro R9700 supported?

Why: This was a general "bug" report from a user who found that the R9700 was unstable or would drop back to CPU mode. This PR provides the formal support this user was asking for.

🔄 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/14021 **Author:** [@yalbik](https://github.com/yalbik) **Created:** 2/2/2026 **Status:** ❌ Closed **Base:** `main` ← **Head:** `ROCm-7` --- ### 📝 Commits (1) - [`136b3bb`](https://github.com/ollama/ollama/commit/136b3bbda4db9bfae0e217b28199872e2cc5848f) Add RDNA4/ROCm-7 support. Tested on Radeon AI Pro R9700 in Windows. ### 📊 Changes **4 files changed** (+99 additions, -10 deletions) <details> <summary>View changed files</summary> 📝 `discover/runner.go` (+42 -1) 📝 `docs/gpu.mdx` (+16 -6) 📝 `envconfig/config.go` (+2 -0) 📝 `llm/server.go` (+39 -3) </details> ### 📄 Description ** REQUIRES ROCm 7.1 on build agents ** This PR adds native discovery and build support for the RDNA 4 architecture (gfx1201). Currently, RDNA 4 cards fail discovery due to a timeout in the HIP handshake and a lack of architecture whitelisting. It was developed with the assistance of Claude Opus 4.5. Changes: * Added gfx1201 to the supported AMD GPU list in gpu/amd_common.go. * Updated CMakePresets.json to include gfx1201 in AMDGPU_TARGETS. * Resolved a library pathing issue where TensileLibrary_lazy_gfx1201.dat was not being correctly located in the rocblas subdirectory. Testing: Verified on Windows 11 with a Radeon AI Pro R9700 and ROCm 7.1.1. Model offloading is now successful. Primary Issues Addressed: [ollama/ollama#13236](https://github.com/ollama/ollama/issues/13236): ROCm GPU discovery times out on AMD Radeon AI PRO R9700 (gfx1201). Why: Discovery logic hangs for 30 seconds and then falls back to 0 layers/CPU. This fix resolves this by whitelisting the architecture. [ollama/ollama#12908](https://github.com/ollama/ollama/issues/12908): ROCM Library not found, wrong location for gfx1201. Why: This addresses the pathing error where the system looks in hipblaslt instead of rocblas for the RDNA 4 .dat files. [ollama/ollama#13908](https://github.com/ollama/ollama/issues/13908): RDNA4 support. Why: This is a broader, more recent community request for native RDNA 4 support on Windows and Linux without using the HSA_OVERRIDE hack. Related/Indirect Issues: [ollama/ollama#13085](https://github.com/ollama/ollama/issues/13085): Is AMD's AI Pro R9700 supported? Why: This was a general "bug" report from a user who found that the R9700 was unstable or would drop back to CPU mode. This PR provides the formal support this user was asking for. --- <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 2026-04-13 00:55:12 -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#14473
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?