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 #14374] RTX 5090 D v2(Blackwell SM_120): MMQ CUDA kernel crash - device kernel image is invalid (v0.16.3, Windows) #9342

New Issue
Open
opened 2026-04-12 22:12:28 -05:00 by GiteaMirror · 3 comments
GiteaMirror commented 2026-04-12 22:12:28 -05:00
Owner
Copy Link

Originally created by @XHXIAIEIN on GitHub (Feb 23, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/14374

What is the issue?

On Windows with RTX 5090 D v2 (China-market variant, GB202-240 die, Blackwell compute capability 12.0), every quantized model fails to load with a CUDA kernel crash in the MMQ (quantized matrix multiplication) function. This affects ALL models (tested qwen3:32b, qwen2.5:14b, qwen2.5:7b, glm-4.7-flash, glm-ocr, deepseek-r1:32b).

The CUDA backend does report SM_120 in its architecture list (CUDA.0.ARCHS=...1200), but the compiled MMQ kernels crash at runtime.

Note: The RTX 5090 D v2 is a China-specific variant using the GB202-240 die (not GB202-300 used by the standard RTX 5090). It has the same 21,760 CUDA cores and compute capability 12.0, but with reduced memory (24 GB / 384-bit vs 32 GB / 512-bit). This variant may not be in the Ollama/llama.cpp CUDA test matrix.

GPU Hardware Details

Spec Standard RTX 5090 RTX 5090 D v2 (this card)
GPU Die GB202-300 GB202-240
GPU Part Number — 2B8C-240-A1
VRAM 32 GB GDDR7 / 512-bit 24 GB GDDR7 / 384-bit
AI TOPS 3,352 2,375
CUDA Cores 21,760 21,760
Compute Capability 12.0 12.0
VBIOS — 98.02.69.00.55

Error

CUDA error: device kernel image is invalid
  current device: 0, in function ggml_cuda_mul_mat_q at C:\a\ollama\ollama\ml\backend\ggml\ggml\src\ggml-cuda\mmq.cu:128
  cudaGetLastError()
C:\a\ollama\ollama\ml\backend\ggml\ggml\src\ggml-cuda\ggml-cuda.cu:94: CUDA error

Full Server Log

load_backend: loaded CPU backend from D:\Agent\Ollama\lib\ollama\ggml-cpu-icelake.dll
ggml_cuda_init: GGML_CUDA_FORCE_MMQ:    no
ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no
ggml_cuda_init: found 1 CUDA devices:
  Device 0: NVIDIA GeForce RTX 5090 D v2, compute capability 12.0, VMM: yes
load_backend: loaded CUDA backend from D:\Agent\Ollama\lib\ollama\cuda_v12\ggml-cuda.dll
system: CUDA.0.ARCHS=500,520,600,610,700,750,800,860,890,900,1200

CUDA error: device kernel image is invalid
  current device: 0, in function ggml_cuda_mul_mat_q at mmq.cu:128

nvidia-smi output:
  Product Name        : NVIDIA GeForce RTX 5090 D v2
  Product Architecture: Blackwell
  GPU Part Number     : 2B8C-240-A1
  VBIOS Version       : 98.02.69.00.55
  Driver Version      : 591.86
  CUDA Version        : 13.1

Steps to Reproduce

  1. Install Ollama 0.16.3 on Windows with an RTX 5090 D v2
  2. Pull a small quantized model: ollama pull qwen2.5:0.5b (~400 MB)
  3. Try to run: ollama run qwen2.5:0.5b "hello"
  4. Model fails to load with the CUDA error above

Minimal reproduction script (Python):

import json, urllib.request

# Requires Ollama running on localhost:11434
payload = json.dumps({
    "model": "qwen2.5:0.5b",  # Any quantized model triggers this
    "prompt": "Hello",
    "stream": False,
    "options": {"num_predict": 1}
}).encode()

req = urllib.request.Request(
    "http://localhost:11434/api/generate",
    data=payload,
    headers={"Content-Type": "application/json"},
)

with urllib.request.urlopen(req, timeout=120) as r:
    data = json.loads(r.read())
    print(data)
    # Expected: response with generated text
    # Actual: {"error": "model failed to load..."} or HTTP 500

Environment

  • OS: Windows 11 Pro 10.0.26200
  • Ollama: 0.16.3
  • GPU: NVIDIA GeForce RTX 5090 D v2 (GB202-240, compute capability 12.0, Blackwell)
  • GPU Part Number: 2B8C-240-A1
  • Driver: 591.86
  • CUDA: 13.1
  • RAM: 64 GB
  • VRAM: 24 GB

What I've Tried

Workaround Result
OLLAMA_FLASH_ATTENTION=false Flash attention disabled, but MMQ crash persists (different code path)
GGML_CUDA_FORCE_CUBLAS=1 as env var No effect — this is a compile-time CMake define, not a runtime env var
OLLAMA_NUM_GPU=0 via Git Bash Not inherited by Ollama runner process on Windows
CUDA_VISIBLE_DEVICES=-1 via batch file Works — falls back to CPU, but loses all GPU acceleration
Reinstall Ollama 0.16.3 Same binary, same issue

Root Cause Analysis

Based on ggml-org/llama.cpp#18331, this is likely caused by an nvcc compiler optimization bug where the CUDA Toolkit generates incorrect machine code for Blackwell SM_120 MMQ kernels at -O3. The pre-built ggml-cuda.dll shipped with Ollama 0.16.3 contains these buggy kernels.

Additionally, since this is the China-variant GB202-240 die (not the standard GB202-300), it may have subtle microarchitectural differences that exacerbate or differently trigger the nvcc compiler bug. This variant may not be included in the Ollama/llama.cpp CUDA test matrix.

The upstream workaround is to either:

  • Compile with -DCMAKE_CUDA_ARCHITECTURES="89" (Ada PTX fallback, works on Blackwell via forward compatibility)
  • Compile with -DCMAKE_CUDA_FLAGS="-O2" (reduced optimization avoids the nvcc bug)
  • Compile with -DGGML_CUDA_FORCE_CUBLAS=ON (bypass MMQ entirely, use cuBLAS)

Expected Behavior

Models should load and run on RTX 5090 D v2 GPU. The CUDA architecture 1200 is listed in the compiled backends, so SM_120 is intended to be supported.

Suggested Fix

  1. Ship the CUDA backend DLL compiled with -DCMAKE_CUDA_ARCHITECTURES="89-real;120" and -O2 flag for sm_120
  2. Or include a runtime fallback to cuBLAS when MMQ kernels fail on Blackwell
  3. Ensure GB202-240 (RTX 5090 D v2) is included in the GPU compatibility test matrix alongside GB202-300 (standard RTX 5090)
Originally created by @XHXIAIEIN on GitHub (Feb 23, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/14374 ## What is the issue? On Windows with **RTX 5090 D v2** (China-market variant, **GB202-240** die, Blackwell compute capability 12.0), **every** quantized model fails to load with a CUDA kernel crash in the MMQ (quantized matrix multiplication) function. This affects ALL models (tested qwen3:32b, qwen2.5:14b, qwen2.5:7b, glm-4.7-flash, glm-ocr, deepseek-r1:32b). The CUDA backend **does** report SM_120 in its architecture list (`CUDA.0.ARCHS=...1200`), but the compiled MMQ kernels crash at runtime. > **Note:** The RTX 5090 D v2 is a China-specific variant using the **GB202-240** die (not GB202-300 used by the standard RTX 5090). It has the same 21,760 CUDA cores and compute capability 12.0, but with reduced memory (24 GB / 384-bit vs 32 GB / 512-bit). This variant may not be in the Ollama/llama.cpp CUDA test matrix. ## GPU Hardware Details | Spec | Standard RTX 5090 | RTX 5090 D v2 (this card) | |---|---|---| | GPU Die | GB202-300 | **GB202-240** | | GPU Part Number | — | **2B8C-240-A1** | | VRAM | 32 GB GDDR7 / 512-bit | **24 GB GDDR7 / 384-bit** | | AI TOPS | 3,352 | **2,375** | | CUDA Cores | 21,760 | 21,760 | | Compute Capability | 12.0 | 12.0 | | VBIOS | — | 98.02.69.00.55 | ## Error ``` CUDA error: device kernel image is invalid current device: 0, in function ggml_cuda_mul_mat_q at C:\a\ollama\ollama\ml\backend\ggml\ggml\src\ggml-cuda\mmq.cu:128 cudaGetLastError() C:\a\ollama\ollama\ml\backend\ggml\ggml\src\ggml-cuda\ggml-cuda.cu:94: CUDA error ``` ## Full Server Log ``` load_backend: loaded CPU backend from D:\Agent\Ollama\lib\ollama\ggml-cpu-icelake.dll ggml_cuda_init: GGML_CUDA_FORCE_MMQ: no ggml_cuda_init: GGML_CUDA_FORCE_CUBLAS: no ggml_cuda_init: found 1 CUDA devices: Device 0: NVIDIA GeForce RTX 5090 D v2, compute capability 12.0, VMM: yes load_backend: loaded CUDA backend from D:\Agent\Ollama\lib\ollama\cuda_v12\ggml-cuda.dll system: CUDA.0.ARCHS=500,520,600,610,700,750,800,860,890,900,1200 CUDA error: device kernel image is invalid current device: 0, in function ggml_cuda_mul_mat_q at mmq.cu:128 ``` ``` nvidia-smi output: Product Name : NVIDIA GeForce RTX 5090 D v2 Product Architecture: Blackwell GPU Part Number : 2B8C-240-A1 VBIOS Version : 98.02.69.00.55 Driver Version : 591.86 CUDA Version : 13.1 ``` ## Steps to Reproduce 1. Install Ollama 0.16.3 on Windows with an RTX 5090 D v2 2. Pull a small quantized model: `ollama pull qwen2.5:0.5b` (~400 MB) 3. Try to run: `ollama run qwen2.5:0.5b "hello"` 4. Model fails to load with the CUDA error above **Minimal reproduction script (Python):** ```python import json, urllib.request # Requires Ollama running on localhost:11434 payload = json.dumps({ "model": "qwen2.5:0.5b", # Any quantized model triggers this "prompt": "Hello", "stream": False, "options": {"num_predict": 1} }).encode() req = urllib.request.Request( "http://localhost:11434/api/generate", data=payload, headers={"Content-Type": "application/json"}, ) with urllib.request.urlopen(req, timeout=120) as r: data = json.loads(r.read()) print(data) # Expected: response with generated text # Actual: {"error": "model failed to load..."} or HTTP 500 ``` ## Environment - **OS**: Windows 11 Pro 10.0.26200 - **Ollama**: 0.16.3 - **GPU**: NVIDIA GeForce RTX 5090 D v2 (GB202-240, compute capability 12.0, Blackwell) - **GPU Part Number**: 2B8C-240-A1 - **Driver**: 591.86 - **CUDA**: 13.1 - **RAM**: 64 GB - **VRAM**: 24 GB ## What I've Tried | Workaround | Result | |---|---| | `OLLAMA_FLASH_ATTENTION=false` | Flash attention disabled, but MMQ crash persists (different code path) | | `GGML_CUDA_FORCE_CUBLAS=1` as env var | **No effect** — this is a compile-time CMake define, not a runtime env var | | `OLLAMA_NUM_GPU=0` via Git Bash | Not inherited by Ollama runner process on Windows | | `CUDA_VISIBLE_DEVICES=-1` via batch file | **Works** — falls back to CPU, but loses all GPU acceleration | | Reinstall Ollama 0.16.3 | Same binary, same issue | ## Root Cause Analysis Based on [ggml-org/llama.cpp#18331](https://github.com/ggml-org/llama.cpp/issues/18331), this is likely caused by an **nvcc compiler optimization bug** where the CUDA Toolkit generates incorrect machine code for Blackwell SM_120 MMQ kernels at `-O3`. The pre-built `ggml-cuda.dll` shipped with Ollama 0.16.3 contains these buggy kernels. Additionally, since this is the **China-variant GB202-240 die** (not the standard GB202-300), it may have subtle microarchitectural differences that exacerbate or differently trigger the nvcc compiler bug. This variant may not be included in the Ollama/llama.cpp CUDA test matrix. The upstream workaround is to either: - Compile with `-DCMAKE_CUDA_ARCHITECTURES="89"` (Ada PTX fallback, works on Blackwell via forward compatibility) - Compile with `-DCMAKE_CUDA_FLAGS="-O2"` (reduced optimization avoids the nvcc bug) - Compile with `-DGGML_CUDA_FORCE_CUBLAS=ON` (bypass MMQ entirely, use cuBLAS) ## Expected Behavior Models should load and run on RTX 5090 D v2 GPU. The CUDA architecture 1200 is listed in the compiled backends, so SM_120 is intended to be supported. ## Suggested Fix 1. Ship the CUDA backend DLL compiled with `-DCMAKE_CUDA_ARCHITECTURES="89-real;120"` and `-O2` flag for sm_120 2. Or include a runtime fallback to cuBLAS when MMQ kernels fail on Blackwell 3. Ensure GB202-240 (RTX 5090 D v2) is included in the GPU compatibility test matrix alongside GB202-300 (standard RTX 5090)
GiteaMirror commented 2026-04-12 22:12:29 -05:00
Author
Owner
Copy Link

@XHXIAIEIN commented on GitHub (Feb 24, 2026):

Update: GPU Hardware Confirmed Working via Alternative Path

Since Ollama's MMQ kernel crashes on this GPU, I switched to HuggingFace Transformers for local inference as a temporary alternative. This confirms the GPU hardware itself is fully functional:

What Works (PyTorch + HuggingFace Transformers)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

# Load Qwen3-8B via HuggingFace (CPU → GPU manual transfer)
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-8B", dtype=torch.bfloat16, low_cpu_mem_usage=True
)
model = model.to("cuda")  # Works on RTX 5090 D v2
Metric Value
Model Qwen3-8B (bf16, ~16 GB)
Runtime PyTorch 2.12.0.dev (nightly, cu128)
VRAM 15.3 GB / 24 GB
Speed ~8.9 tok/s

Successfully ran a complete RAG pipeline (vector retrieval + generation + self-reflection) with correct, high-confidence results.

What This Tells Us About the Ollama Issue

  • The GPU's CUDA compute (sm_120) works correctly — matrix operations, weight loading, bf16 inference all pass
  • The problem is specifically in llama.cpp's compiled MMQ CUDA kernels — the nvcc-generated machine code for sm_120 is faulty (as noted in the upstream ggml-org/llama.cpp#18331)
  • cuBLAS path works on this GPU (PyTorch uses cuBLAS internally), which supports the suggestion that compiling with -DGGML_CUDA_FORCE_CUBLAS=ON would fix it

Temporary Alternative for RTX 5090 D v2 Users

If you need local LLM inference while waiting for the Ollama fix:

  1. Install PyTorch nightly with cu128: pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128
  2. Use HuggingFace Transformers to load models directly (bypass device_map="auto" — that has a separate bug)
  3. Models up to ~16 GB bf16 fit in the 24 GB VRAM

This is obviously not a fix — just confirming the hardware is fine and the issue is purely in the MMQ kernel compilation.

<!-- gh-comment-id:3950122718 --> @XHXIAIEIN commented on GitHub (Feb 24, 2026): ## Update: GPU Hardware Confirmed Working via Alternative Path Since Ollama's MMQ kernel crashes on this GPU, I switched to **HuggingFace Transformers** for local inference as a temporary alternative. This confirms the GPU hardware itself is fully functional: ### What Works (PyTorch + HuggingFace Transformers) ```python import torch from transformers import AutoModelForCausalLM, AutoTokenizer # Load Qwen3-8B via HuggingFace (CPU → GPU manual transfer) model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-8B", dtype=torch.bfloat16, low_cpu_mem_usage=True ) model = model.to("cuda") # Works on RTX 5090 D v2 ``` | Metric | Value | |---|---| | Model | Qwen3-8B (bf16, ~16 GB) | | Runtime | PyTorch 2.12.0.dev (nightly, cu128) | | VRAM | 15.3 GB / 24 GB | | Speed | ~8.9 tok/s | Successfully ran a complete RAG pipeline (vector retrieval + generation + self-reflection) with correct, high-confidence results. ### What This Tells Us About the Ollama Issue - **The GPU's CUDA compute (sm_120) works correctly** — matrix operations, weight loading, bf16 inference all pass - **The problem is specifically in llama.cpp's compiled MMQ CUDA kernels** — the nvcc-generated machine code for sm_120 is faulty (as noted in the upstream [ggml-org/llama.cpp#18331](https://github.com/ggml-org/llama.cpp/issues/18331)) - **cuBLAS path works on this GPU** (PyTorch uses cuBLAS internally), which supports the suggestion that compiling with `-DGGML_CUDA_FORCE_CUBLAS=ON` would fix it ### Temporary Alternative for RTX 5090 D v2 Users If you need local LLM inference while waiting for the Ollama fix: 1. Install PyTorch nightly with cu128: `pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu128` 2. Use HuggingFace Transformers to load models directly (bypass `device_map="auto"` — that has a [separate bug](https://github.com/huggingface/accelerate/issues/3933)) 3. Models up to ~16 GB bf16 fit in the 24 GB VRAM This is obviously not a fix — just confirming the hardware is fine and the issue is purely in the MMQ kernel compilation.
GiteaMirror commented 2026-04-12 22:12:30 -05:00
Author
Owner
Copy Link

@XHXIAIEIN commented on GitHub (Feb 24, 2026):

Correction: AI TOPS Specification

The hardware comparison table in the original report contains an error:

Field Originally reported for RTX 5090 D v2 Correction
AI TOPS 2,375 This figure is from the original RTX 5090 D (GB202-250), not the D v2 (GB202-240). The D v2's actual AI TOPS are not independently confirmed — the 384-bit / 1,344 GB/s bandwidth reduction (vs. 512-bit / 1,792 GB/s on the original D) may further reduce effective AI throughput beyond the firmware-limited 2,375 TOPS.

All other hardware specs (die GB202-240, 21,760 CUDA cores, 24 GB / 384-bit, sm_120) are confirmed correct via nvidia-smi -q and PyTorch device properties (170 SMs × 128 = 21,760 cores, compute capability 12.0).

The core issue (MMQ kernel crash) and root cause analysis remain unchanged — this is an nvcc compiler bug for sm_120 MMQ kernels, not a hardware issue.

<!-- gh-comment-id:3951108439 --> @XHXIAIEIN commented on GitHub (Feb 24, 2026): ## Correction: AI TOPS Specification The hardware comparison table in the original report contains an error: | Field | Originally reported for RTX 5090 D v2 | Correction | |---|---|---| | AI TOPS | 2,375 | **This figure is from the original RTX 5090 D (GB202-250), not the D v2 (GB202-240).** The D v2's actual AI TOPS are not independently confirmed — the 384-bit / 1,344 GB/s bandwidth reduction (vs. 512-bit / 1,792 GB/s on the original D) may further reduce effective AI throughput beyond the firmware-limited 2,375 TOPS. | All other hardware specs (die GB202-240, 21,760 CUDA cores, 24 GB / 384-bit, sm_120) are confirmed correct via `nvidia-smi -q` and PyTorch device properties (170 SMs × 128 = 21,760 cores, compute capability 12.0). The core issue (MMQ kernel crash) and root cause analysis remain unchanged — this is an nvcc compiler bug for sm_120 MMQ kernels, not a hardware issue.
GiteaMirror commented 2026-04-12 22:12:30 -05:00
Author
Owner
Copy Link

@lingfan36 commented on GitHub (Feb 25, 2026):

👋 你好！

看起来你是安装相关的问题。我们整理了 Ollama 安装常见问题的解决方案：

🔗 安装问题解决方案: https://ollamahub.space/pages/solutions/installation/

如果这里没有你遇到的问题，欢迎在下方补充详细信息（如操作系统、错误信息等），我们会及时更新文档。

由 OllamaHub 自动生成

<!-- gh-comment-id:3957146782 --> @lingfan36 commented on GitHub (Feb 25, 2026): 👋 你好！ 看起来你是安装相关的问题。我们整理了 Ollama 安装常见问题的解决方案： 🔗 **安装问题解决方案**: https://ollamahub.space/pages/solutions/installation/ 如果这里没有你遇到的问题，欢迎在下方补充详细信息（如操作系统、错误信息等），我们会及时更新文档。 --- *由 OllamaHub 自动生成*
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
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#9342
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?