github-issues/T-017C.md

ISSUE T-017C: RX 6700 XT Optimization and Tuning Framework

Status: status:complete Parent: #139 (GENESIS) GitHub ID: 175

Summary

Owns the hardware-specific optimization contract for the workstation.

Completed

• Forced HSA_OVERRIDE_GFX_VERSION=10.3.0 for RDNA 2 stability.
• Implemented PYTORCH_HIP_ALLOC_CONF for memory fragmentation control.
• Defined MatFormer strategy for elastic inference (800M -> 200M).
• Integrated 4-bit (NF4) quantization plan.

New Sub-tasks / Refinements

• Implement VRAM spike monitoring in Trackio logs.
• Create benchmark script for fp16 vs int8 vs nf4 on RX 6700 XT.
• Tuning: verify HSA_ENABLE_SDMA=0 impact on large model transfers.