[deepseek-v4] Bump sgl-deep-gemm for packed FP4×FP4 mega-MoE (W4A4)

[Romaosir]

Bumps the pinned sgl-deep-gemm version (pyproject + docker ARG) to the release carrying the new sm100_fp4_mega_moe packed FP4×FP4 (W4A4) mega-MoE kernel, and propagates the selection.

In _apply_mega_moe_dg_env, when SGLANG_OPT_DEEPGEMM_MEGA_MOE_USE_FP4_ACTS=1 → os.environ.setdefault("DG_MEGA_MOE_FP4", "1") (same setdefault pattern as the existing flags; an external DG_* override still wins). With this, the existing W4A4 config transparently routes to the packed kernel — no other code changes.

Depends on

• [sm100] Packed FP4×FP4 mega-MoE kernel (W4A4) with per-band BLOCK_K DeepGEMM#44 ([sm100] Packed FP4×FP4 mega-MoE kernel (W4A4) with per-band BLOCK_K DeepGEMM#44) — [sm100] Packed FP4×FP4 mega-MoE kernel (W4A4) with per-band BLOCK_K — and its subsequent sgl-deep-gemm release. The pinned version string here should be aligned to that release tag before merge.

Measured impact (4× B200, DSV4-Flash, TP4+DP4)

Routing the existing W4A4 config to the packed kernel yields, end-to-end at production concurrency (c ≥ 128): prefill input-throughput +2–7% vs the 0.1.0 W4A4 kernel (binding-controlled), mixed throughput +2%, median TPOT down to −11% at c=512. Accuracy preserved (GSM8K 0.966 / HumanEval 0.898, ≥ W4A8 baseline). Full kernel-level detail in the paired DeepGEMM PR sgl-project/DeepGEMM#44 (sgl-project/DeepGEMM#44); the measured data this routing change unlocks is reproduced below.

---

Appendix: measured data (the kernel this routing change selects)

All on 4× B200, DSV4-Flash (hidden=4096 / inter=2048 / E=256 / topk=6), TP=4 + DP=4. W4A4-Ours = the packed sm100_fp4_mega_moe selected by this PR's DG_MEGA_MOE_FP4; W4A4-old = the existing 0.1.0-release W4A4 kernel (sgl-project/DeepGEMM#27) — both measured through the same DeepGEMM binding (DG_FORCE_FP8FP4 toggle) so the comparison isolates the kernel from packaging; W4A8 = the FP8-acts reference (0.1.0 release).

A. End-to-end 5-concurrency sweep

Sequential, same GPU group, one server at a time; 1 measured rep/cell after a per-concurrency warmup; num_prompts raised for steady state.

Prefill (in=1024, out=1) — input throughput (tok/s) · median TTFT (ms):

conc	Ours tput	old tput	W4A8 tput	Δ tput Ours/old	Ours TTFT	old TTFT	Δ TTFT Ours/old
1	5847	5886	5683	−0.7%	156.8	154.8	+1.4%
4	14802	14478	10095	+2.2%	289.5	292.2	−0.9%
32	54114	54755	52762	−1.2%	417.6	402.2	+3.8%
128	126557	117838	123033	+7.4%	1110.6	1145.8	−3.1%
512	133116	130299	124708	+2.2%	4575.7	4672.6	−2.1%

Mixed (in=1024, out=256) — total throughput (tok/s) · median TPOT (ms):

conc	Ours tput	old tput	W4A8 tput	Δ tput Ours/old	Ours TPOT	old TPOT	Δ TPOT Ours/old
1	342	338	340	+1.1%	14.02	14.17	−1.1%
4	1268	1291	1274	−1.8%	14.55	14.82	−1.8%
32	8212	8055	8149	+2.0%	17.82	18.11	−1.6%
128	24982	24544	24204	+1.8%	22.40	22.63	−1.0%
512	51295	50201	50363	+2.2%	33.50	37.81	−11.4%

The win surfaces at production concurrency (c ≥ 128) where mega-MoE's share is largest; low concurrency is within the e2e noise band (decode is ~82% the shared sm100_tf32_hc_prenorm_gemm, ~8% mega-MoE).

B. NCU Speed-of-Light (1-rank isolation, binding-controlled, ntok=512)

metric	W4A4-Ours (packed, bk256)	W4A4-old (fp8_fp4, bk128)	W4A8 (fp8 acts)
Duration (µs)	537.6	712.5	753.2
Compute (SM) %	33.6	28.6	42.5
Memory %	85.2	64.3	61.4
DRAM throughput (TB/s)	6.54	4.93	4.71
L2 %	56.1	42.4	40.3
Achieved occupancy %	22.5	22.4	22.5
Registers/thread	128	128	128

Ours +24.6% vs W4A4-old, +28.6% vs W4A8 (1-rank). Same occupancy/regs → the gain is the data path: block_k=256 sustains 85% memory SOL / 6.54 TB/s DRAM vs W4A4-old's 64% / 4.93. 1-rank over-weights the memory regime; the e2e sweep above is the realistic magnitude.

C. Accuracy (eval_v2, DSV4-Flash)

eval	W4A4-Ours	W4A4-old	W4A8
GSM8K	0.966	0.962	0.962
HumanEval (pass@1)	0.898	0.888	0.887

At or above both baselines on every eval — the routing change preserves accuracy.

[gemini-code-assist]

Code Review

This pull request updates the sgl-deep-gemm dependency from version 0.1.0 to 0.1.3 in the Dockerfile and Python dependencies. It enables routing to DeepGEMM's packed FP4xFP4 mega-MoE kernel (sm100_fp4_mega_moe) by setting the DG_MEGA_MOE_FP4 environment variable when FP4 activations are enabled. Additionally, the documentation and code comments have been updated to reflect this new behavior. There are no review comments to address, and I have no further feedback to provide.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.