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 #14439] [Windows] ROCm 7.1 / HIP Error on Radeon RX 9000 Series (gfx1200) - CUBLAS_STATUS_INTERNAL_ERROR #35136

New Issue
Closed
opened 2026-04-22 19:25:01 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-04-22 19:25:01 -05:00
Owner
Copy Link

Originally created by @diegoabreu15 on GitHub (Feb 26, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14439

Describe the bug
On Windows 11, using an AMD Radeon RX 9060 XT (architecture gfx1200) with ROCm 7.1 / HIP SDK, Ollama fails to offload models entirely to the GPU. Even for small models like Gemma 3:4b (4.3 GB), the system performs a partial offload (approx. 82% CPU / 18% GPU) or crashes with a connection error.

Log Snippets
When running with OLLAMA_DEBUG=1, the following error occurs:
wsarecv: An existing connection was forcibly closed by the remote host.
Internal HIP errors indicate CUBLAS_STATUS_INTERNAL_ERROR or memory allocation failures despite having 8GB of VRAM available.

Workaround discovered
Setting OLLAMA_VULKAN=1 and HIP_VISIBLE_DEVICES=-1 stabilizes the model and allows 100% GPU offload, but loses the performance benefits of the native ROCm 7.1 path.

Hardware & Software

OS: Windows 11

GPU: AMD Radeon RX 9060 XT (8GB VRAM)

RAM: 32GB

Driver: AMD Software: Adrenalin Edition (Latest)

ROCm/HIP Version: 7.1

Ollama Version: Latest

Relevant log output

time=2026-02-25T18:43:00.828-03:00 level=INFO source=routes.go:1768 msg="vram-based default context" total_vram="8.0 GiB" default_num_ctx=4096

[GIN] 2026/02/25 - 18:45:47 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:45:47 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:45:47 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:45:47 | 200 |      10.906ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:45:48 | 401 |    199.6413ms |       127.0.0.1 | POST     "/api/me"

[GIN] 2026/02/25 - 18:45:48 | 401 |    166.6388ms |       127.0.0.1 | POST     "/api/me"

[GIN] 2026/02/25 - 18:45:48 | 404 |      1.0193ms |       127.0.0.1 | POST     "/api/show"

[GIN] 2026/02/25 - 18:45:52 | 404 |      1.0417ms |       127.0.0.1 | POST     "/api/show"

[GIN] 2026/02/25 - 18:45:52 | 200 |      1.0347ms |       127.0.0.1 | GET      "/api/tags"

time=2026-02-25T18:45:53.377-03:00 level=INFO source=download.go:179 msg="downloading e7b273f96360 in 16 862 MB part(s)"

[GIN] 2026/02/25 - 18:46:22 | 200 |       1.537ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:46:52 | 200 |      1.0207ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:47:22 | 200 |      2.5673ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:47:52 | 200 |       518.9µs |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:48:22 | 200 |      1.0197ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:48:52 | 200 |      1.5398ms |       127.0.0.1 | GET      "/api/tags"

time=2026-02-25T18:49:05.530-03:00 level=INFO source=download.go:179 msg="downloading fa6710a93d78 in 1 7.2 KB part(s)"

time=2026-02-25T18:49:06.903-03:00 level=INFO source=download.go:179 msg="downloading f60356777647 in 1 11 KB part(s)"

time=2026-02-25T18:49:08.285-03:00 level=INFO source=download.go:179 msg="downloading d8ba2f9a17b3 in 1 18 B part(s)"

time=2026-02-25T18:49:09.685-03:00 level=INFO source=download.go:179 msg="downloading 776beb3adb23 in 1 489 B part(s)"

[GIN] 2026/02/25 - 18:49:22 | 200 |       516.8µs |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:49:26 | 200 |         3m33s |       127.0.0.1 | POST     "/api/pull"

[GIN] 2026/02/25 - 18:49:26 | 200 |      7.2178ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:49:26 | 200 |    213.0554ms |       127.0.0.1 | POST     "/api/show"

[GIN] 2026/02/25 - 18:49:26 | 200 |    230.9763ms |       127.0.0.1 | POST     "/api/show"

time=2026-02-25T18:49:26.467-03:00 level=INFO source=server.go:431 msg="starting runner" cmd="C:\\Users\\pc765xt\\AppData\\Local\\Programs\\Ollama\\ollama.exe runner --ollama-engine --port 53122"

time=2026-02-25T18:49:26.978-03:00 level=INFO source=cpu_windows.go:148 msg=packages count=1

time=2026-02-25T18:49:26.978-03:00 level=INFO source=cpu_windows.go:195 msg="" package=0 cores=6 efficiency=0 threads=12

time=2026-02-25T18:49:27.105-03:00 level=INFO source=server.go:247 msg="enabling flash attention"

time=2026-02-25T18:49:27.107-03:00 level=INFO source=server.go:431 msg="starting runner" cmd="C:\\Users\\pc765xt\\AppData\\Local\\Programs\\Ollama\\ollama.exe runner --ollama-engine --model C:\\Users\\pc765xt\\.ollama\\models\\blobs\\sha256-e7b273f9636059a689e3ddcab3716e4f65abe0143ac978e46673ad0e52d09efb --port 53132"

time=2026-02-25T18:49:27.110-03:00 level=INFO source=sched.go:491 msg="system memory" total="31.9 GiB" free="22.8 GiB" free_swap="21.0 GiB"

time=2026-02-25T18:49:27.110-03:00 level=INFO source=sched.go:498 msg="gpu memory" id=0 library=ROCm available="5.9 GiB" free="6.3 GiB" minimum="457.0 MiB" overhead="0 B"

time=2026-02-25T18:49:27.110-03:00 level=INFO source=server.go:757 msg="loading model" "model layers"=25 requested=-1

time=2026-02-25T18:49:27.139-03:00 level=INFO source=runner.go:1411 msg="starting ollama engine"

time=2026-02-25T18:49:27.147-03:00 level=INFO source=runner.go:1446 msg="Server listening on 127.0.0.1:53132"

time=2026-02-25T18:49:27.153-03:00 level=INFO source=runner.go:1284 msg=load request="{Operation:fit LoraPath:[] Parallel:1 BatchSize:512 FlashAttention:Enabled KvSize:4096 KvCacheType: NumThreads:6 GPULayers:25[ID:0 Layers:25(0..24)] MultiUserCache:false ProjectorPath: MainGPU:0 UseMmap:false}"

time=2026-02-25T18:49:27.201-03:00 level=INFO source=ggml.go:136 msg="" architecture=gptoss file_type=MXFP4 name="" description="" num_tensors=459 num_key_values=32

load_backend: loaded CPU backend from C:\Users\pc765xt\AppData\Local\Programs\Ollama\lib\ollama\ggml-cpu-haswell.dll

ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no

ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no

ggml_cuda_init: found 1 ROCm devices:

  Device 0: AMD Radeon RX 9060 XT, gfx1200 (0x1200), VMM: no, Wave Size: 32, ID: 0

load_backend: loaded ROCm backend from C:\Users\pc765xt\AppData\Local\Programs\Ollama\lib\ollama\rocm\ggml-hip.dll

time=2026-02-25T18:49:27.245-03:00 level=INFO source=ggml.go:104 msg=system CPU.0.SSE3=1 CPU.0.SSSE3=1 CPU.0.AVX=1 CPU.0.AVX2=1 CPU.0.F16C=1 CPU.0.FMA=1 CPU.0.BMI2=1 CPU.0.LLAMAFILE=1 CPU.1.LLAMAFILE=1 ROCm.0.NO_VMM=1 ROCm.0.NO_PEER_COPY=1 ROCm.0.PEER_MAX_BATCH_SIZE=128 compiler=cgo(clang)

HIP Library Path: C:\WINDOWS\SYSTEM32\amdhip64_7.dll

ROCm error: CUBLAS_STATUS_INTERNAL_ERROR

  current device: 0, in function ggml_cuda_op_mul_mat_cublas at C:/a/ollama/ollama/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu:1415

  hipblasGemmEx(ctx.cublas_handle(id), HIPBLAS_OP_T, HIPBLAS_OP_N, row_diff, src1_ncols, ne10, &alpha_f32, src0_ptr, HIPBLAS_R_16B, ne00, src1_ptr, HIPBLAS_R_16B, ne10, &beta_f32, dst_bf16.get(), HIPBLAS_R_16B, ldc, HIPBLAS_R_32F, HIPBLAS_GEMM_DEFAULT)

C:/a/ollama/ollama/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error

[GIN] 2026/02/25 - 18:49:54 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:49:54 | 200 |      2.2667ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:49:59 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:49:59 | 200 |      2.1903ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:50:00 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:50:00 | 200 |      2.0663ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:50:00 | 200 |            0s |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:50:00 | 200 |      3.2627ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:50:06 | 200 |       516.6µs |       127.0.0.1 | GET      "/api/version"

[GIN] 2026/02/25 - 18:50:06 | 200 |      2.6283ms |       127.0.0.1 | GET      "/api/tags"

[GIN] 2026/02/25 - 18:50:36 | 200 |      1.0187ms |       127.0.0.1 | GET      "/api/tags"

OS

Windows

GPU

AMD

CPU

AMD

Ollama version

0.17.0

Originally created by @diegoabreu15 on GitHub (Feb 26, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14439 Describe the bug On Windows 11, using an AMD Radeon RX 9060 XT (architecture gfx1200) with ROCm 7.1 / HIP SDK, Ollama fails to offload models entirely to the GPU. Even for small models like Gemma 3:4b (4.3 GB), the system performs a partial offload (approx. 82% CPU / 18% GPU) or crashes with a connection error. Log Snippets When running with OLLAMA_DEBUG=1, the following error occurs: wsarecv: An existing connection was forcibly closed by the remote host. Internal HIP errors indicate CUBLAS_STATUS_INTERNAL_ERROR or memory allocation failures despite having 8GB of VRAM available. Workaround discovered Setting OLLAMA_VULKAN=1 and HIP_VISIBLE_DEVICES=-1 stabilizes the model and allows 100% GPU offload, but loses the performance benefits of the native ROCm 7.1 path. Hardware & Software OS: Windows 11 GPU: AMD Radeon RX 9060 XT (8GB VRAM) RAM: 32GB Driver: AMD Software: Adrenalin Edition (Latest) ROCm/HIP Version: 7.1 Ollama Version: Latest ### Relevant log output ```shell time=2026-02-25T18:43:00.828-03:00 level=INFO source=routes.go:1768 msg="vram-based default context" total_vram="8.0 GiB" default_num_ctx=4096 [GIN] 2026/02/25 - 18:45:47 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:45:47 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:45:47 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:45:47 | 200 | 10.906ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:45:48 | 401 | 199.6413ms | 127.0.0.1 | POST "/api/me" [GIN] 2026/02/25 - 18:45:48 | 401 | 166.6388ms | 127.0.0.1 | POST "/api/me" [GIN] 2026/02/25 - 18:45:48 | 404 | 1.0193ms | 127.0.0.1 | POST "/api/show" [GIN] 2026/02/25 - 18:45:52 | 404 | 1.0417ms | 127.0.0.1 | POST "/api/show" [GIN] 2026/02/25 - 18:45:52 | 200 | 1.0347ms | 127.0.0.1 | GET "/api/tags" time=2026-02-25T18:45:53.377-03:00 level=INFO source=download.go:179 msg="downloading e7b273f96360 in 16 862 MB part(s)" [GIN] 2026/02/25 - 18:46:22 | 200 | 1.537ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:46:52 | 200 | 1.0207ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:47:22 | 200 | 2.5673ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:47:52 | 200 | 518.9µs | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:48:22 | 200 | 1.0197ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:48:52 | 200 | 1.5398ms | 127.0.0.1 | GET "/api/tags" time=2026-02-25T18:49:05.530-03:00 level=INFO source=download.go:179 msg="downloading fa6710a93d78 in 1 7.2 KB part(s)" time=2026-02-25T18:49:06.903-03:00 level=INFO source=download.go:179 msg="downloading f60356777647 in 1 11 KB part(s)" time=2026-02-25T18:49:08.285-03:00 level=INFO source=download.go:179 msg="downloading d8ba2f9a17b3 in 1 18 B part(s)" time=2026-02-25T18:49:09.685-03:00 level=INFO source=download.go:179 msg="downloading 776beb3adb23 in 1 489 B part(s)" [GIN] 2026/02/25 - 18:49:22 | 200 | 516.8µs | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:49:26 | 200 | 3m33s | 127.0.0.1 | POST "/api/pull" [GIN] 2026/02/25 - 18:49:26 | 200 | 7.2178ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:49:26 | 200 | 213.0554ms | 127.0.0.1 | POST "/api/show" [GIN] 2026/02/25 - 18:49:26 | 200 | 230.9763ms | 127.0.0.1 | POST "/api/show" time=2026-02-25T18:49:26.467-03:00 level=INFO source=server.go:431 msg="starting runner" cmd="C:\\Users\\pc765xt\\AppData\\Local\\Programs\\Ollama\\ollama.exe runner --ollama-engine --port 53122" time=2026-02-25T18:49:26.978-03:00 level=INFO source=cpu_windows.go:148 msg=packages count=1 time=2026-02-25T18:49:26.978-03:00 level=INFO source=cpu_windows.go:195 msg="" package=0 cores=6 efficiency=0 threads=12 time=2026-02-25T18:49:27.105-03:00 level=INFO source=server.go:247 msg="enabling flash attention" time=2026-02-25T18:49:27.107-03:00 level=INFO source=server.go:431 msg="starting runner" cmd="C:\\Users\\pc765xt\\AppData\\Local\\Programs\\Ollama\\ollama.exe runner --ollama-engine --model C:\\Users\\pc765xt\\.ollama\\models\\blobs\\sha256-e7b273f9636059a689e3ddcab3716e4f65abe0143ac978e46673ad0e52d09efb --port 53132" time=2026-02-25T18:49:27.110-03:00 level=INFO source=sched.go:491 msg="system memory" total="31.9 GiB" free="22.8 GiB" free_swap="21.0 GiB" time=2026-02-25T18:49:27.110-03:00 level=INFO source=sched.go:498 msg="gpu memory" id=0 library=ROCm available="5.9 GiB" free="6.3 GiB" minimum="457.0 MiB" overhead="0 B" time=2026-02-25T18:49:27.110-03:00 level=INFO source=server.go:757 msg="loading model" "model layers"=25 requested=-1 time=2026-02-25T18:49:27.139-03:00 level=INFO source=runner.go:1411 msg="starting ollama engine" time=2026-02-25T18:49:27.147-03:00 level=INFO source=runner.go:1446 msg="Server listening on 127.0.0.1:53132" time=2026-02-25T18:49:27.153-03:00 level=INFO source=runner.go:1284 msg=load request="{Operation:fit LoraPath:[] Parallel:1 BatchSize:512 FlashAttention:Enabled KvSize:4096 KvCacheType: NumThreads:6 GPULayers:25[ID:0 Layers:25(0..24)] MultiUserCache:false ProjectorPath: MainGPU:0 UseMmap:false}" time=2026-02-25T18:49:27.201-03:00 level=INFO source=ggml.go:136 msg="" architecture=gptoss file_type=MXFP4 name="" description="" num_tensors=459 num_key_values=32 load_backend: loaded CPU backend from C:\Users\pc765xt\AppData\Local\Programs\Ollama\lib\ollama\ggml-cpu-haswell.dll ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ggml_cuda_init: found 1 ROCm devices: Device 0: AMD Radeon RX 9060 XT, gfx1200 (0x1200), VMM: no, Wave Size: 32, ID: 0 load_backend: loaded ROCm backend from C:\Users\pc765xt\AppData\Local\Programs\Ollama\lib\ollama\rocm\ggml-hip.dll time=2026-02-25T18:49:27.245-03:00 level=INFO source=ggml.go:104 msg=system CPU.0.SSE3=1 CPU.0.SSSE3=1 CPU.0.AVX=1 CPU.0.AVX2=1 CPU.0.F16C=1 CPU.0.FMA=1 CPU.0.BMI2=1 CPU.0.LLAMAFILE=1 CPU.1.LLAMAFILE=1 ROCm.0.NO_VMM=1 ROCm.0.NO_PEER_COPY=1 ROCm.0.PEER_MAX_BATCH_SIZE=128 compiler=cgo(clang) HIP Library Path: C:\WINDOWS\SYSTEM32\amdhip64_7.dll ROCm error: CUBLAS_STATUS_INTERNAL_ERROR current device: 0, in function ggml_cuda_op_mul_mat_cublas at C:/a/ollama/ollama/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu:1415 hipblasGemmEx(ctx.cublas_handle(id), HIPBLAS_OP_T, HIPBLAS_OP_N, row_diff, src1_ncols, ne10, &alpha_f32, src0_ptr, HIPBLAS_R_16B, ne00, src1_ptr, HIPBLAS_R_16B, ne10, &beta_f32, dst_bf16.get(), HIPBLAS_R_16B, ldc, HIPBLAS_R_32F, HIPBLAS_GEMM_DEFAULT) C:/a/ollama/ollama/ml/backend/ggml/ggml/src/ggml-cuda/ggml-cuda.cu:94: ROCm error [GIN] 2026/02/25 - 18:49:54 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:49:54 | 200 | 2.2667ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:49:59 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:49:59 | 200 | 2.1903ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:50:00 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:50:00 | 200 | 2.0663ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:50:00 | 200 | 0s | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:50:00 | 200 | 3.2627ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:50:06 | 200 | 516.6µs | 127.0.0.1 | GET "/api/version" [GIN] 2026/02/25 - 18:50:06 | 200 | 2.6283ms | 127.0.0.1 | GET "/api/tags" [GIN] 2026/02/25 - 18:50:36 | 200 | 1.0187ms | 127.0.0.1 | GET "/api/tags" ``` ### OS Windows ### GPU AMD ### CPU AMD ### Ollama version 0.17.0
GiteaMirror added the bug label 2026-04-22 19:25:01 -05:00
GiteaMirror commented 2026-04-22 19:25:03 -05:00
Author
Owner
Copy Link

@diegoabreu15 commented on GitHub (Feb 26, 2026):

rocm.gfx1200.for.rocm.6.2.4-no-optimized.zip

Simply replace these files in the ollama lib and rockblas folders.

<!-- gh-comment-id:3968581510 --> @diegoabreu15 commented on GitHub (Feb 26, 2026): [rocm.gfx1200.for.rocm.6.2.4-no-optimized.zip](https://github.com/user-attachments/files/25585247/rocm.gfx1200.for.rocm.6.2.4-no-optimized.zip) Simply replace these files in the ollama lib and rockblas folders.
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
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#35136
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?