kaggle(rebake): hide the GPU from torch — 8 backends died to the P100 sm_60 drop

v2 got past the source_model blocker (only the 4 deliberately-blank entries
still report it) and hit a new uniform failure. The log names it:

    torch.AcceleratorError: CUDA error: no kernel image is available for
    execution on the device                                            x8

That is kaggle_usage.md gotcha #23 verbatim — Kaggle's preinstalled torch has
dropped sm_60, and gotcha #21 says P100 is effectively the only draw. nemotron,
canary-1b-v2, parakeet-tdt-0.6b-ja, granite-4.1-plus/-nar, voxtral4b-realtime
and vibevoice-asr each downloaded their model, loaded it, then died on the first
kernel launch — ~100 s of wasted download apiece.

MY FAULT, and the fix is to undo my own choice: I set enable_gpu:"true" in this
kernel's metadata. The canonical regression kernel is deliberately
enable_gpu:"false" because reference dumps do not need a GPU. Enabling one
bought nothing and exposed every torch dumper to the lottery.

Not simply flipping the flag, though: a CPU-only Kaggle worker loses internet
(gotcha #3) and this job must pull models from HF. So the kernel keeps its GPU
allocation and hides it from torch with CUDA_VISIBLE_DEVICES="". Same effect,
network retained.

REMAINING, separate cause, not addressed here: a dependency tail — funasr,
mimo_audio_tokenizer, whisper, qwen_asr, fireredasr all ModuleNotFoundError, and
transformers too old for VoxtralRealtimeForConditionalGeneration. Those need
per-backend pip installs in the canonical suite, which is a wider change.

Log saved to /mnt/volume1/tmp-overflow/rebake-v2.log BEFORE re-pushing this
time: kernels_logs serves only the current version, so v1's log was destroyed by
the v2 push along with the tracebacks I needed.

Author: crispasr integration

tools/kaggle/regression-rebake/rebake.py

@@ -25,7 +25,22 @@
 os.environ["CRISPASR_REGRESSION_UPLOAD"] = "1"
 os.environ["CRISPASR_REGRESSION_BUILD"] = os.environ.get("CRISPASR_REGRESSION_BUILD", "cpu")
 
-SCRIPT_VERSION = "2026-09-07-rebake-1"
+# HIDE THE GPU FROM TORCH. v2 lost 8 backends to
+#   torch.AcceleratorError: CUDA error: no kernel image is available
+# which is kaggle_usage.md gotcha #23: Kaggle's preinstalled torch has dropped
+# sm_60, so a P100 draw is fatal to any torch code — and P100 is effectively the
+# only draw right now (#21). nemotron, canary-1b-v2, parakeet-tdt-0.6b-ja,
+# granite-4.1-*, voxtral4b and vibevoice each downloaded their model, loaded it,
+# and died on the first kernel launch, ~100 s each.
+#
+# Reference dumps do not need a GPU — the canonical regression kernel is
+# deliberately enable_gpu:false. But a CPU-only Kaggle worker loses internet
+# (gotcha #3) and this job must pull from HF, so the kernel keeps its GPU and
+# simply does not show it to torch. Same effect, without giving up the network.
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+os.environ["CRISPASR_REF_DEVICE"] = "cpu"
+
+SCRIPT_VERSION = "2026-09-07-rebake-3-cpu-torch"
 WORK = pathlib.Path("/kaggle/working")
 
 # Clone into a SEPARATE bootstrap dir: the canonical script manages its own