about summary refs log tree commit diff
diff options
context:
space:
mode:
-rw-r--r--README.md16
-rw-r--r--session.md32
2 files changed, 42 insertions, 6 deletions
diff --git a/README.md b/README.md
index 4b01029..a759e92 100644
--- a/README.md
+++ b/README.md
@@ -281,11 +281,17 @@ The scores are meaningless biologically (the circos PNG is a stub), but
 every stage of the DJL pipeline exercises: `libtorch` dlopen, TorchScript
 deserialisation, forward pass, output translation.
 
-The GPU leg needs a DJL runtime with kernels for the installed card.
-CUDA 12.1 (the newest DJL native published for PyTorch 2.2.2) has no
-`sm_120` images, so RTX 50-series (Blackwell) cards fail with *"no kernel
-image is available for execution on the device"*; either wait for a cu128
-runtime or point `PYTORCH_LIBRARY_PATH` at a system libtorch.
+The GPU leg needs a DJL runtime with kernels for the installed card, and
+no DJL release has them for RTX 50-series (Blackwell, `sm_120`): the
+newest native DJL publishes is `pytorch-native-cu124` 2.5.1, i.e. CUDA
+12.4, whose `libtorch_cuda.so` carries `sm_35` … `sm_90` kernels only
+(checked against the 2.2.2/cu121 bundle here).  Blackwell needs CUDA 12.8+
+(PyTorch 2.7+), so the card fails with *"no kernel image is available for
+execution on the device"*.  The plausible route is a newer DJL
+(`pytorch-jni` 2.7.1 exists for DJL 0.34–0.39, and every DJL class v-chord
+uses is unchanged in 0.39) plus a system libtorch 2.7.1+cu128 pointed at
+by `PYTORCH_LIBRARY_PATH` — that has to be tried on a host where the GPU
+device is actually visible.
 
 ## Related
 
diff --git a/session.md b/session.md
index d18fe00..f4522d7 100644
--- a/session.md
+++ b/session.md
@@ -199,8 +199,38 @@ ravanan (`cwl-directory-and-array-support`):
 
 | item                                     | state |
 |------------------------------------------|-------|
+| GPU (RTX 5080) from this sandbox         | impossible: `/dev` is a private tmpfs with no `nvidia*` nodes, `CapEff=0` so `mknod` is denied, and no `libcuda.so*` exists anywhere (`/`, `/gnu/store`); only the kernel module is visible via `/proc/driver/nvidia` |
+| GPU with the packaged DJL stack          | impossible even on the host: newest DJL native is `pytorch-native-cu124` 2.5.1 (CUDA 12.4 -> `sm_90`), the card is `sm_120`; our cu121 `libtorch_cuda.so` lists `sm_35`…`sm_90` only |
+| GPU, plausible route (host only)         | DJL 0.39 (`pytorch-jni` 2.7.1, all 17 DJL classes used are unchanged) + system libtorch 2.7.1+cu128 via `PYTORCH_LIBRARY_PATH`; needs the device visible and a ~2.5 GB download |
 | `workflows/wigits.cwl` execution         | not run end to end: needs an Ensembl cache, GC profile, BAF loci VCF, LILAC resources and indexed BAMs; `qc.cwl` and `dir-handoff.cwl` are the executable proofs |
 | ravanan cwltest e2e / conformance suites | not buildable here: `cwltest` -> `python-cheroot` fails two SSL assertions on this OpenSSL |
-| v-chord GPU leg                          | needs a DJL runtime with `sm_120` kernels (cu128) for RTX 50-series cards |
 | ravanan branch                           | `cwl-directory-and-array-support`, 6 commits local; not merged, not sent upstream |
 
+## 11. GPU investigation
+
+The host has an NVIDIA GeForce RTX 5080 (driver 610.43.02, open kernel
+module), visible through `/proc/driver/nvidia`.  Two independent blockers
+prevent using it:
+
+| blocker                                        | evidence |
+|------------------------------------------------|----------|
+| no device access from the sandbox              | `/dev` is a private tmpfs carrying only `null`, `zero`, `full`, `random`, `urandom`, `tty`, `fuse`; `CapEff=0` so `mknod` fails with `EPERM`; no `libcuda.so*` under `/`, `/gnu/store` or `/usr/lib` |
+| the card needs CUDA 12.8+, DJL ships CUDA 12.4 | newest DJL artifact is `pytorch-native-cu124` 2.5.1; `strings libtorch_cuda.so` from the 2.2.2/cu121 bundle lists `sm_35`…`sm_90` and no `sm_100`/`sm_120`, so a Blackwell card fails with "no kernel image is available for execution on the device" |
+
+What would be needed to go further, in order of cost:
+
+1. run `test/test-vchord.sh` on the host, where `/dev/nvidia*` and
+   `/usr/lib/x86_64-linux-gnu/libcuda.so.1` exist — still fails on the
+   second blocker;
+2. bump the channel's DJL stack to 0.39 (the API is compatible: all 17
+   DJL classes v-chord and v-cuppa import still exist), install
+   `pytorch-jni` 2.7.1-0.39.0, and point `PYTORCH_LIBRARY_PATH` at a
+   libtorch 2.7.1+cu128 build from download.pytorch.org (~2.5 GB).  That
+   is the first combination that could execute on this card, and it needs
+   a machine with the device visible;
+3. or bind-mount `/dev/nvidia*` and the host's `libcuda.so.1` into the
+   sandbox — only the host owner can do that, and it does not help until
+   step 2 is done.
+
+Until then the CPU leg of `test/test-vchord.sh` and `v-chord-e2e` in the
+CWL suite remain the GPU-less proofs.