Update SETUP_MIAAI.md: bitsandbytes sm_120 patch, OOM fixes, working training config

This commit is contained in:
2026-05-13 18:58:51 +00:00
parent 8693a1f61b
commit 74f2263ac7

View File

@@ -1,11 +1,13 @@
# Axolotl Setup — miaai (RTX 5080, CUDA 13.2)
## System Info
- GPU: NVIDIA RTX 5080 (16GB VRAM)
- Driver: 580.126.09 — max CUDA 13.0 (nvcc from conda resolves to 13.2)
- OS: Ubuntu (Python 3.13 system — do NOT use system Python for ML)
- GPU: NVIDIA RTX 5080 (16GB VRAM, sm_120 / Blackwell)
- Driver: 580.126.09 — max CUDA 13.0 shown by nvidia-smi, but nvcc from conda is 13.2
- OS: Ubuntu 25.10 (Python 3.13 system — do NOT use system Python for ML)
- Axolotl branch: `activeblue/main`
---
## One-time Setup
### 1. Install Miniconda
@@ -32,13 +34,16 @@ git rebase upstream/main # keeps activeblue patches on top
git push origin activeblue/main --force-with-lease
```
### 4. Install CUDA toolkit (needed to compile flash-attn)
### 4. Install CUDA toolkit (needed to compile flash-attn and bitsandbytes)
```bash
conda install -y -c "nvidia/label/cuda-12.8.0" cuda-toolkit
export CUDA_HOME=$CONDA_PREFIX
export PATH=$CUDA_HOME/bin:$PATH
```
> NOTE: Despite installing from the cuda-12.8.0 channel, conda resolves nvcc to **13.2.78**.
> This is fine — use cu132 everywhere to match.
### 5. Install PyTorch — use cu132 (matches nvcc from conda)
> NOTE: torchaudio has no cu132 wheel — skip it, not needed for LLM training
```bash
@@ -52,26 +57,109 @@ pip install -e "."
```
> **flash-attn compiles CUDA kernels from source — takes 1525 min on 10 cores of i7-14700K.**
> Always set `MAX_JOBS` to the number of available CPU cores to parallelize and speed up compilation:
> Always set `MAX_JOBS` to the number of available CPU cores:
```bash
MAX_JOBS=10 pip install flash-attn --no-build-isolation
```
### 7. Compile bitsandbytes from source for sm_120 (RTX 5080 / Blackwell)
The prebuilt bitsandbytes wheels do not include sm_120 support and CUDA 13.2 dropped sm_5053.
You must compile from source with a patched CMakeLists.txt.
```bash
# Clone bitsandbytes v0.49.1
git clone --branch v0.49.1 --depth 1 https://github.com/bitsandbytes-foundation/bitsandbytes.git /tmp/bnb_0491
cd /tmp/bnb_0491
# Patch CMakeLists.txt: override arch list to sm_120 only, just before the foreach loop
# (cmake >= 3.23.0 skips the manual arch block and uses its own built-in list which lacks sm_120)
sed -i '/ foreach(capability \${CMAKE_CUDA_ARCHITECTURES_ALL})/i\ # RTX 5080 sm_120 patch: override before capability list is built\n set(CMAKE_CUDA_ARCHITECTURES_ALL 120)' CMakeLists.txt
# Verify the patch landed at the right line
grep -n "ARCHITECTURES_ALL\|foreach" CMakeLists.txt | tail -5
# Should show: set(CMAKE_CUDA_ARCHITECTURES_ALL 120) immediately before the foreach line
# Configure — must point cmake at conda's nvcc
cmake \
-DCMAKE_CUDA_COMPILER=/opt/miniconda3/envs/axolotl/bin/nvcc \
-DCOMPUTE_BACKEND=cuda \
-S /tmp/bnb_0491 \
-B /tmp/bnb_0491/build 2>&1 | grep -E "(Capabilit|CUDA Ver|Error)"
# Expected: "CUDA Capabilities Selected: 120"
# Build (j10 uses 10 cores — adjust to your CPU)
cmake --build /tmp/bnb_0491/build -j10
# Install into conda site-packages
SITE_PKG=/opt/miniconda3/envs/axolotl/lib/python3.11/site-packages
cp -r /tmp/bnb_0491/bitsandbytes "$SITE_PKG/"
# Verify
python3 -c "
import torch, bitsandbytes as bnb
x = torch.randn(64, 64, device='cuda')
l = bnb.nn.Linear8bitLt(64, 64).cuda()
print('bitsandbytes CUDA OK:', l(x).shape)
"
```
---
## Every Session (after first-time setup)
```bash
export PATH="/opt/miniconda3/bin:$PATH"
conda activate axolotl
export CUDA_HOME=$CONDA_PREFIX
export PATH=$CUDA_HOME/bin:$PATH
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
cd /home/tocmo0nlord/axolotl
```
---
## Training Config — human_chat_qlora.yml
Key settings that work on RTX 5080 (16GB):
| Setting | Value | Notes |
|---|---|---|
| `sequence_len` | `2048` | 4096 causes OOM during loss computation (logits x 128k vocab) |
| `micro_batch_size` | `1` | Keep low; effective batch = micro x grad_accum |
| `gradient_accumulation_steps` | `8` | Effective batch = 8 |
| `adapter` | `qlora` | QLoRA 4-bit via bitsandbytes |
| `attn_implementation` | `flash_attention_2` | Not the deprecated `flash_attention: true` |
| `type` (datasets) | `chat_template` | Not the deprecated `sharegpt` |
Dataset fields for SlimOrca / OpenHermes-2.5 (sharegpt-format with different field names):
```yaml
datasets:
- path: Open-Orca/SlimOrca
type: chat_template
field_messages: conversations
message_field_role: from
message_field_content: value
split: "train[:3%]"
```
## Run Training
```bash
axolotl train human_chat_qlora.yml
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
axolotl train ~/human_chat_qlora.yml
```
Expected startup sequence:
1. Config validation + capability detection (shows `sm_120`)
2. Dataset tokenization (~65k samples, ~30 seconds)
3. `Loading weights: 100% 291/291`
4. `trainable params: 167,772,160 || all params: 8,198,033,408 || trainable%: 2.05`
5. Initial eval: loss ~0.81, perplexity ~2.25, VRAM ~8.5GB
6. Training steps at ~2.6 it/s, VRAM ~9-10GB
---
## Common Pitfalls Encountered
| Problem | Cause | Fix |
|---|---|---|
| `externally-managed-environment` | System Python 3.13 blocks pip | Use conda env, never system pip |
@@ -81,3 +169,10 @@ axolotl train human_chat_qlora.yml
| `torchaudio` not found for cu132 | No cu132 wheel exists | Skip torchaudio — not needed |
| `src refspec main does not match` | Fork default branch is `activeblue/main` | `git push origin activeblue/main` |
| flash-attn compile is slow | Single-threaded by default | Set `MAX_JOBS=<cpu_count>` before pip install |
| `nvcc fatal: Unsupported gpu architecture 'compute_50'` | bitsandbytes CMakeLists.txt hardcodes sm_50; CUDA 13.2 dropped it | Patch CMakeLists.txt (see step 7 above) |
| `CUDA Capabilities Selected: 50;52;...` (ignores sm_120) | cmake >= 3.23 built-in arch list lacks sm_120 | Add `set(CMAKE_CUDA_ARCHITECTURES_ALL 120)` before foreach loop |
| `BackendUnavailable: scikit_build_core` | pip install of bnb tries to rebuild | Copy .so directly to site-packages instead |
| `torch.OutOfMemoryError` during eval | logits tensor (batch x 4096 x 128k vocab) too large | Set `sequence_len: 2048`, `micro_batch_size: 1` |
| `type: sharegpt` deprecation warning | axolotl removed sharegpt type | Use `type: chat_template` with field mappings |
| `flash_attention: true` deprecation | Old config key removed | Use `attn_implementation: flash_attention_2` |
| Capybara dataset `field_messages null` | Capybara uses input/output format, not conversations | Switch to SlimOrca or OpenHermes-2.5 |