…4074)

Add GLM-5.2 MXFP8 (per_1x32) tuned/untuned MoE configs and merge
best-of-both GLM-5.2 MXFP4 tuning into glm5_fp4_tuned_fmoe.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…FP4 MoE (#4095)

* Add per-token tuned configs for GLM-5.2-FP8 fp8 GEMM and MoE

Adds per-token quant tuned (and source untuned) configs for GLM-5.2-FP8
on gfx950 (TP4), complementing the block-scale (per_1x128) configs. Same
M/N/K shapes as the block-scale set, quantized per-token instead:

- a8w8_bpreshuffle GEMM: 4 shapes (N6144/K3072, N6144/K4096,
  N6144/K6144, N7168/K512), swept over M in {1..32768}.
- fused MoE (QuantType.per_Token): hidden_dim=6144, inter_dim=512,
  expert=257, topk=9, swept over tokens {1..32768}.

Both quant families coexist in the model_configs glob merge; the runtime
matches rows by q_type, so per-token and per_1x128 configs do not
collide. No code change required. Tuning completed with zero failures.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Add tuned ptpc a8w8_bpreshuffle GEMM configs for GLM-5.2 N2624/K6144 and N4096/K2048

These two fp8 GEMM shapes were previously missing from the ptpc enumeration,
causing the server to fall back to default (untuned) configs. Tuned all 16 M
values {1..32768} for each on gfx950 (256 CU).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* Drop glm5.2 ptpc GEMM rows duplicating existing tuned configs

Remove 37 shapes from the glm5.2 a8w8_bpreshuffle tuned/untuned files whose
(gfx,cu_num,M,N,K,q_dtype_w) already exist in other merged configs on main
(minimax_m3: N6144/K3072 + N6144/K6144; global default: 5x N7168/K512).
This keeps existing config files untouched and avoids the runtime merge
auto-dedup error. glm5.2 retains only shapes unique to it (59 rows).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

* [fmoe][gfx950] Add GLM-5.2 MXFP8/MXFP4 moe tune config (cherry-pick #4074)

Add GLM-5.2 MXFP8 (per_1x32) tuned/untuned MoE configs and merge
best-of-both GLM-5.2 MXFP4 tuning into glm5_fp4_tuned_fmoe.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* [fmoe][gfx950] Drop glm5.2 fp8 fmoe configs; keep only flydsl-over-flydsl in glm5_fp4

Per review:
1. Remove glm5.2_fp8_{tuned,untuned}_fmoe.csv (no longer needed).
2. In glm5_fp4_tuned_fmoe.csv, only let the merged GLM-5.2 MXFP4 rows override
   main when both sides are FlyDSL. Revert the 5 shapes where the incoming row
   used a CK (moe_ck2stages) kernel over main's FlyDSL row (inter256 token1/2,
   inter512 token4/16384/32768) back to main's FlyDSL config. The remaining 10
   overrides are FlyDSL-over-FlyDSL and are kept.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* [fmoe] Rename GLM-5.2 MXFP8 fmoe configs to <model>_<quant>_tuned_fmoe convention

mxfp8_{tuned,untuned}_fmoe_glm5_2.csv -> glm5_2_mxfp8_{tuned,untuned}_fmoe.csv
to match the model-first naming used by every other fmoe config
(e.g. glm5_fp4_tuned_fmoe.csv, minimax_m3_mxfp8_tuned_fmoe.csv). Content is
unchanged; the new names still contain the tuned_fmoe/untuned_fmoe substring
the config auto-merge globs on, and runtime selection is by shape-key content
not filename, so config lookup is unaffected.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* [fmoe] Rename GLM-5.2 MXFP8 fmoe configs glm5_2 -> glm5

glm5_2_mxfp8_{tuned,untuned}_fmoe.csv -> glm5_mxfp8_{tuned,untuned}_fmoe.csv,
matching the glm5 prefix used by glm5_fp4_tuned_fmoe.csv. Pure rename; content
unchanged and the tuned_fmoe/untuned_fmoe substring is preserved so config
lookup is unaffected.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* [Fix] exclude untuned_fmoe CSVs from CK 2stages prebuild enumeration

The prebuild variant enumerator globs "*tuned_fmoe*.csv", which also
matches "*untuned_fmoe*.csv" as a substring. Untuned CSVs describe
problem shapes to tune, not resolved kernels, and carry no kernelName
columns, so the flydsl-skip guard does not exclude them. Their rows then
fall through to CK 2stages codegen.

GLM-5.2 MXFP8 (Silu, per_1x32, f8/f8) is the first untuned combo that
also slips past the existing swiglu-skip, forcing prebuild to AOT-compile
a CK 2stages instance that has no working implementation:

  module_moe_ck2stages_f8_f8_preshuffle_on_b16_silu_per_1x32_mulWeightStage2
  threadwise_tensor_slice_transfer_v7r3_scatter.hpp: no matching call to 'unpack2'

Filter out any path containing "untuned" so untuned CSVs never feed the
prebuild enumerator. MXFP8 MoE is served by FlyDSL kernels (the tuned CSV
rows carry flydsl_* kernel names and are already skipped).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: zejunchen-zejun <zejun.chen@amd.com>