perf(rocm): enable Liger fused RMSNorm/SwiGLU for AMD DFlash configs

[zhentaocc]

Summary

Turns on Liger fused RMSNorm/SwiGLU (model.use_liger_kernel: true) for the two AMD ROCm Qwen3.5-4B DFlash example configs. Validated as a clean win on MI355X (gfx950) with real specforge train, not a profiler estimate.

• examples/configs/offline/colocated/qwen3.5-4b-dflash-offline-amd.yaml
• examples/configs/online/disaggregated/external/qwen3.5-4b-dflash-online-amd.yaml

Measured on MI355X (Qwen3.5-4B DFlash, offline, batch_size=2, S=2048, num_anchors=512)

Method: container based on sglang-v0.5.14 (torch 2.9.1+rocm7.2.0), single card, max_steps=200, discard first 20 warmup steps, steady-state mean of perf/global_samples_per_second (n=19); peak VRAM polled across all 8 cards (others idle). Same seed, loss checked step-by-step.

	tokens/s (samples/s)	s/step	peak VRAM	loss @ step200
use_liger_kernel=false	12.68	0.623	35.37 GiB	7.0105
use_liger_kernel=true	13.97	0.565	30.18 GiB	7.0101

+10.2% throughput, −5.2 GiB / −15% peak memory, loss parity.

Note: the CUDA-profiler-estimated FLCE (fused Linear+CE) tier was dropped — it is numerically correct but 2.5× slower on gfx950 (Liger's chunked recompute overhead outweighs the logits memory-bandwidth savings). Fused AdamW was also measured (+0.9%, within noise) and dropped. Only RMSNorm/SwiGLU is enabled here.

Depends on #773

This PR does not include the flex-backend fix; it must be merged on top of #773 (or after #773 lands). On torch<2.11, use_liger_kernel=true drives flex-attention through the inductor autotune lowering path, where the unpatched kernel_options={\"BACKEND\":\"TRITON\"} leaks a bare TRITON identifier → NameError: 'TRITON' is not defined. #773 gates that on torch>=2.11. Merging this config change before #773 would break the AMD configs on ROCm torch<2.11.

Test plan

• Loss parity vs use_liger_kernel=false (same seed, per-step)
• MI355X real-train throughput + peak-mem A/B
• Rebased onto latest main (config-dir restructure followed; fix(dflash): gate flex_attention BACKEND=TRITON on torch>=2.11 (fixes ROCm) #773 commit removed)