#2413 docs/AMD_HIP.md: short GPU stalls while the kernel compacts memory, and how to keep them out of a speed comparison

本文

## Title
Issue: none (related: #750, #920 and #1705, the reclaim stalls of the section before it; #2126 is a different compaction stall)

## Summary

On a Ryzen AI Max+ 395 (gfx1151), a prompt that met heavy page migration was usually 0.1-1.6 s slower, mostly the first prompt
after a start. That can hide a 1-2% change: in one 11-pair 4K A/B, the off / on ratio of `prompt_ms` had a 95% interval of
[0.979, 1.078] on each engine's first prompt and [1.016, 1.020] on the second. This adds a short section after the reclaim
section: what was seen, and how to measure when the difference is small enough for these stalls to hide it.

## What changed

- `docs/AMD_HIP.md`: one new section between the reclaim section and "RDNA4 (gfx1201)". No code changes.

Measured on gfx1151, 128 GB, ROCm 7.14.1, Linux 7.0, UD-IQ4_XS with the experts read from the GGUF in place (engine 7695859a;
the later prompts on a0116b27):

| prompt | met more than 50,000 migrated pages | of those, 0.1 s or more slower |
|---|---:|---:|
| first 4K prompt after a fresh start | 11 of 50 | 10, by 0.17-1.60 s (than the second prompt) |
| second 4K prompt | 0 of 50 | - |
| later 4K-128K prompts of an A/B | 3 of 138 | 2, by 0.12 and 0.27 s (than the median of the same kind) |

- KFD queue evictions, read as root through `AMDKFD_IOC_SMI_EVENTS` over 20 starts: in all 6 prompts that met any page
  migration, in none of the 34 that did not.
- `GPU_PINNED_MIN_XFER_SIZE=1048576`, `HSA_USERPTR_FOR_PAGED_MEM=0` and the two compaction sysctls each left slow first
  prompts (worst 0.39, 1.27 and 0.25 s in 10 starts each).
- Still seen on 296b59d2.

The runs, the KFD event log and a script that recomputes these numbers:
https://github.com/libratechw/Strata/tree/a13f5285516c3f39090a8705103084d94a9fecb7

## Extra Notes

- One machine, one kernel and ROCm version. Whether the restore goes through the `-EAGAIN` retries of the section above was not
  traced.

🤖 Generated with [Claude Code](https://claude.com/claude-code)