Feature/954 update rocm

[lauri9]

Bump ROCm to version 10.1.0 (pre-release).

Closes #954

Summary of changes

Bump ROCm stack.

Installation of amd-smi has changed in rocm-systems, and this is updated in the Dockerfile.

Two workarounds are needed:

• libsqlite3-dev is installed to workaround rocprofiler-sdk SQLite header gap - [Issue]: rocprofiler-sdk SQLite include path relies on incidental manylinux header ROCm/TheRock#5868
• Install tcl to work around profiler-hub SQLite amalgamation build failure - [Issue] rocprofiler-systems configure fails on build hosts without Tcl: profiler-hub's bundled SQLite3 amalgamation requires tclsh ROCm/rocm-systems#10059

Tasks

• Bump ROCm
• Update MIOpen tuning databases
• Verify performance and visual quality on gfx950
• Verify performance and visual quality on gfx942

Performance

MI350X

model	ROCm bump	baseline main	Ratio baseline to best
flux.usp	0.79	0.85	108.0%
flux.single_gpu	2.36	2.41	102.4%
flux.usp_2k	2.20	2.25	102.0%
flux.single_gpu_2k	11.93	12.16	101.9%
flux2.default	4.51	4.52	100.3%
flux2.quantgemm.gfx950	2.47	2.47	100.1%
flux2.quantgemm_sageattn.gfx950	2.57	2.61	101.4%
flux2.ti2i_2k	12.54	12.53	100.0%
flux2.t2i_1k	2.93	2.91	99.5%
flux2.t2i_2k	10.68	10.72	100.4%
flux2.t-multi-i2i_1k	5.93	5.95	100.3%
flux2.t-multi-i2i_2k	15.04	15.02	99.9%
flux2_klein.default	0.48	0.49	100.7%
flux_kontext.default	1.50	1.50	100.0%
hunyuanvideo.default	102.63	104.06	101.4%
hunyuanvideo.fp8attn	65.21	66.39	101.8%
hunyuanvideo.sageattn.gfx950	87.27	88.00	100.8%
hunyuanvideo_1_5.default	117.00	116.71	99.8%
hunyuanvideo_1_5.distilled.gfx950	55.43	55.18	99.6%
hunyuanvideo_1_5.distilled_sparse.gfx950	46.33	46.04	99.4%
ltx2_3.default	15.75	16.17	102.7%
qwen_image.default	4.92	4.89	99.4%
qwen_image_edit.default	15.02	15.12	100.7%
stablediffusion_3_5.default	0.81	0.81	99.5%
wan2_1.default	75.08	74.54	99.3%
wan2_2.default	75.94	75.99	100.1%
wan2_2.quantgemm.gfx950	69.31	68.76	99.2%
wan2_2.quantgemm_fp8attn.gfx950	41.53	41.17	99.1%
wan2_2.quantgemm_sageattn.gfx950	54.94	54.72	99.6%
wan2_2.quantgemm_spargeattn.gfx950	31.30	31.03	99.1%
z_image.default	2.65	2.67	101.0%

MI300X

Verified no regressions, but not sharing a full table here as runs were completed on multiple different nodes hence not easy to compare.

Known issue - profiling

Profiling on diffusion workloads is broken on PyTorch 2.9. When PyTorch profiler enters warmup or active mode, the system stutters and gaps appear in the GPU traces. Runtimes can be up to 4x compared to benchmarking. This issue was tested to be resolved after updating to PyTorch 2.13 on the same ROCm version. The issue likely relates to rocprofiler-sdk updates or other updates lower in the stack. PyTorch 2.9 is unlikely to get a fix to support newer ROCm, as it has already fallen off the support matrix of TheRock.

The issue will be resolved by updating to a newer PyTorch stack in a later PR. Verified to be resolved on PyTorch 2.13.