[oobabooga]

The dequantize_block and convert_unary kernels pass ne01 directly as the CUDA grid y-dimension, but grid.y is limited to 65535. When ne01 exceeds this, the kernel launch fails with cudaErrorInvalidConfiguration.

This happens when using llama-server with flash attention, quantized KV cache, multiple parallel slots, and long context. With multiple slots the KV caches are non-contiguous, so the NC dequantization path is taken, and ne01 (the KV cache length) ends up as grid.y.

The grid.z dimension was already capped at 65535 with a grid-stride loop. This applies the same pattern to grid.y.

Closes #19487
Closes #19724

Reproduction

# Start server with GLM-4.7-Flash, flash attention, quantized KV, 3 slots
llama-server \
    --model GLM-4.7-Flash-UD-Q6_K_XL.gguf \
    --ctx-size 608256 \
    --flash-attn on \
    --cache-type-k q8_0 \
    --cache-type-v q8_0 \
    -np 3 \
    --no-kv-unified \
    -ngl 999 \
    --port 5000

# Generate a ~150k token prompt
python3 -c "
import json
text = 'The quick brown fox jumps over the lazy dog. ' * 15000
req = {'model': 'test', 'messages': [{'role': 'user', 'content': text}], 'max_tokens': 32}
open('/tmp/prompt.json', 'w').write(json.dumps(req))
"

# Send 3 concurrent requests - crashes or produces CUDA errors
curl -X POST http://localhost:5000/v1/chat/completions -H "Content-Type: application/json" -d @/tmp/prompt.json &
curl -X POST http://localhost:5000/v1/chat/completions -H "Content-Type: application/json" -d @/tmp/prompt.json &
curl -X POST http://localhost:5000/v1/chat/completions -H "Content-Type: application/json" -d @/tmp/prompt.json &
wait

# CUDA error: invalid configuration argument

[jeffbolznv]

Is it possible to add backend tests for this?