RTX 5090: incorrect max_threads_per_processor in SUPPORTED_GPUS (1536 should be 1024)

[WinstonFrankBot]

Summary

The SUPPORTED_GPUS entry for the RTX 5090 in api/gpu.py has an incorrect max_threads_per_processor value. This causes all RTX 5090 miners to fail device_info verification immediately.

Related: chutes-miner #135

The Bug

The 5090 entry lists:

"max_threads_per_processor": 1536

The actual hardware reports (via GraVal device info):

max_threads_per_processor: 1024

The device_info_challenge hash fails because 1024 != 1536. Verification rejects in ~234ms — consistent with an immediate hash mismatch.

Evidence

1. RTX PRO 6000 — same Blackwell sm_120 architecture as the 5090 — is correctly listed with max_threads_per_processor: 1024 in the same SUPPORTED_GPUS dict.

2. Our miner's SQL INSERT log during the verification flow shows the device reporting 1024, confirming this is what the hardware actually returns.

3. The 5090 is the only sm_120 GPU listed with 1536. The value appears to be incorrect — possibly confused with CUDA cores per SM or another spec.

Fix

One-line change in api/gpu.py:

# RTX 5090 entry in SUPPORTED_GPUS
"max_threads_per_processor": 1024  # was 1536

Impact

Every RTX 5090 miner on the network fails verification and cannot serve workloads. This has been the case since the 5090 was added to SUPPORTED_GPUS.