Point scprint's triton at its own ptxas

CHANGELOG.md

@@ -57,6 +57,8 @@
 * Fix `methods/scgpt_zeroshot` and `methods/scgpt_finetuned` failing to build: stop installing `flash-attn`. Both
     scripts pass `use_fast_transformer=False`, so it was never used. Behind it sat three more pins with no python 3.12
     wheels, now `numpy<2`, `torchtext==0.17.2` and `transformers==4.36.2`.
+* Fix `methods/scprint` erroring with `Triton only support CUDA 10.0 or higher`: point triton at the ptxas it bundles,
+    since its version table stops at CUDA 12 and the base image now ships CUDA 13.
 * Fix `methods/scgpt_finetuned` erroring on `'csr_matrix' object has no attribute 'A'`: `.A` was removed from
     scipy sparse matrices in scipy 1.14.
 * Fix `methods/geneformer` erroring with a 404 on every dictionary and model file: upstream reorganised the repo for

src/methods/scprint/config.vsh.yaml

@@ -73,6 +73,11 @@ engines:
   - type: docker
     image: openproblems/base_pytorch_nvidia:1
     setup:
+      # scprint pins torch 2.2 / triton 2.2, whose ptx version table stops at CUDA 12, while the base
+      # image now ships CUDA 13 at /usr/local/cuda. Point triton at the ptxas it bundles itself.
+      - type: docker
+        env:
+          - TRITON_PTXAS_PATH=/usr/local/lib/python3.12/dist-packages/triton/third_party/cuda/bin/ptxas
       - type: python
         pip:
           - scprint==2.3.5
@@ -86,7 +91,6 @@ engines:
         script: from scdataloader.utils import populate_my_ontology; populate_my_ontology()
 runners:
   - type: executable
-   # docker_run_args: --gpus all
   - type: nextflow
     directives:
       label: [hightime, highmem, midcpu, gpu, highsharedmem]