[TRTLLM-17033][feat] Qwen3.8-Flash-Next modeling_v2: CUTEDSL, MTP 1..7, video, FP8/NVFP4 KV, bf16 state, canonical QSA selection, breakable prefill graph, GB300 fast paths by Wanli-Jiang · Pull Request #20276 · NVIDIA/TensorRT-LLM

[TRTLLM-17033][feat] Engine, QSA and modeling_v2 catalog support for the Qwen3.8-Flash-Next features

Wanli-Jiang

committed

6d48622

[TRTLLM-17033][feat] Qwen3.8-Flash-Next modeling_v2 target: MTP 1..7, CuteDSL experts, video, FP8/NVFP4 KV, bf16 GDN state, breakable prefill graph

Wanli-Jiang

committed

e10d6c6

[TRTLLM-17033][test] Accuracy gates and test lists for the Qwen3.8-Flash-Next modeling_v2 features

Wanli-Jiang

committed

c8ed8cb

[TRTLLM-16460][perf] Shared kernels for the Qwen3.8-Flash-Next target: low-M bf16 GEMM, FlashInfer GDN entries, PDL QSA index scores

Wanli-Jiang

committed

439b505

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: attention ops

Wanli-Jiang

committed

acafef1

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: Gated-DeltaNet ops and their gated norm

Wanli-Jiang

committed

ef3a736

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: hyper-connection ops

Wanli-Jiang

committed

d7b77d5

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: MoE, GEMM, PLE and elementwise ops

Wanli-Jiang

committed

a14b0b9

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: run the fast paths

Wanli-Jiang

committed

614965c

[TRTLLM-16460][chore] Qwen3.8-Flash-Next target: remove dead paths and duplicated helpers

Wanli-Jiang

committed

6e34769

[TRTLLM-17033][test] Run most qwen4_exp accuracy gates post-merge

Wanli-Jiang

committed

370bcd0

[TRTLLM-17033][fix] Give the beam-search admission stub the model engine the model hook reads

Wanli-Jiang

committed

1f515cf

[TRTLLM-16460][fix] Qwen3.8-Flash-Next target: keep MTP verify steps on the composed attention prologue

Wanli-Jiang

committed

7c9e767

[TRTLLM-16460][fix] Qwen3.8-Flash-Next target: gated norm sends a z that does not cover x to the checked op

Wanli-Jiang

committed

bf7b60f

[TRTLLM-17033][chore] modeling_v2 routing: give each family's route the ModelConfig and LLM args

Wanli-Jiang

committed

66ab699

[TRTLLM-16460][chore] Qwen3.8-Flash-Next target: own low_m_bf16_mm and flashinfer_chunk_gated_delta_rule

Wanli-Jiang

committed

7c8b176

[TRTLLM-17033][test] Qwen3.8-Flash-Next tests: assert the decode overlap, pin int64 positions, run the MoE scalars on CPU

Wanli-Jiang

committed

22c5e87

[TRTLLM-16460][chore] Qwen3.8-Flash-Next target: import the CuTe helpers from main's kernels tree

Wanli-Jiang

committed

37e2407

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: one qsa_topk_expand_decode kernel for every row count

Wanli-Jiang

committed

ebae095

[TRTLLM-17033][fix] thop attention: FP8 context FMHA for a non-MLA NVFP4 KV cache only on SM100f

Wanli-Jiang

committed

65f7c68

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: keep the second expert copy only where its launches can happen

Wanli-Jiang

committed

bf4dbeb

[TRTLLM-17033][fix] QSA cache manager: side-plane transfer roles independent of the KV head split

Wanli-Jiang

committed

b3e6c75

[TRTLLM-17033][test] Qwen3.8-Flash-Next accuracy gates: 26 cases down to 11

Wanli-Jiang

committed

a8ce4e5

[TRTLLM-17033][test] Qwen3.8-Flash-Next unit tests: certify the envelope the target runs

Wanli-Jiang

committed

078261b

[TRTLLM-17033][test] Qwen3.8-Flash-Next accuracy: six cases, each engine running every check that fits

Wanli-Jiang

committed

8e12b66

[TRTLLM-17033][test] Qwen3.8-Flash-Next ops: at most three cells per op, one property and one guard test

Wanli-Jiang

committed

c9288d9

[TRTLLM-17033][test] Qwen3.8-Flash-Next accuracy: acceptance floors for the fp8 and NVFP4 KV engines

Wanli-Jiang

committed

8094d42

[TRTLLM-17033][fix] PyExecutor admission: the model hook and its tests next to main's KV retention check

Wanli-Jiang

committed

257d5f0

Navigation Menu

Repository navigation

[TRTLLM-17033][feat] Qwen3.8-Flash-Next modeling_v2: CUTEDSL, MTP 1..7, video, FP8/NVFP4 KV, bf16 state, canonical QSA selection, breakable prefill graph, GB300 fast paths - #20276

[TRTLLM-17033][feat] Qwen3.8-Flash-Next modeling_v2: CUTEDSL, MTP 1..7, video, FP8/NVFP4 KV, bf16 state, canonical QSA selection, breakable prefill graph, GB300 fast paths#20276
Wanli-Jiang merged 28 commits into
NVIDIA:mainfrom
Wanli-Jiang:user/williamj/qwen38fn-staircase-gaps

Commits on Oct 22, 2026

[TRTLLM-17033][feat] Engine, QSA and modeling_v2 catalog support for the Qwen3.8-Flash-Next features

[TRTLLM-17033][feat] Qwen3.8-Flash-Next modeling_v2 target: MTP 1..7, CuteDSL experts, video, FP8/NVFP4 KV, bf16 GDN state, breakable prefill graph

[TRTLLM-17033][test] Accuracy gates and test lists for the Qwen3.8-Flash-Next modeling_v2 features

[TRTLLM-16460][perf] Shared kernels for the Qwen3.8-Flash-Next target: low-M bf16 GEMM, FlashInfer GDN entries, PDL QSA index scores

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: attention ops

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: Gated-DeltaNet ops and their gated norm

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: hyper-connection ops

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: MoE, GEMM, PLE and elementwise ops

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: run the fast paths

[TRTLLM-16460][chore] Qwen3.8-Flash-Next target: remove dead paths and duplicated helpers

[TRTLLM-17033][test] Run most qwen4_exp accuracy gates post-merge

[TRTLLM-17033][fix] Give the beam-search admission stub the model engine the model hook reads

[TRTLLM-16460][fix] Qwen3.8-Flash-Next target: keep MTP verify steps on the composed attention prologue

[TRTLLM-16460][fix] Qwen3.8-Flash-Next target: gated norm sends a z that does not cover x to the checked op

[TRTLLM-17033][chore] modeling_v2 routing: give each family's route the ModelConfig and LLM args

[TRTLLM-16460][chore] Qwen3.8-Flash-Next target: own low_m_bf16_mm and flashinfer_chunk_gated_delta_rule

[TRTLLM-17033][test] Qwen3.8-Flash-Next tests: assert the decode overlap, pin int64 positions, run the MoE scalars on CPU

[TRTLLM-16460][chore] Qwen3.8-Flash-Next target: import the CuTe helpers from main's kernels tree

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: one qsa_topk_expand_decode kernel for every row count

[TRTLLM-17033][fix] thop attention: FP8 context FMHA for a non-MLA NVFP4 KV cache only on SM100f

[TRTLLM-16460][perf] Qwen3.8-Flash-Next target: keep the second expert copy only where its launches can happen

[TRTLLM-17033][fix] QSA cache manager: side-plane transfer roles independent of the KV head split

[TRTLLM-17033][test] Qwen3.8-Flash-Next accuracy gates: 26 cases down to 11

[TRTLLM-17033][test] Qwen3.8-Flash-Next unit tests: certify the envelope the target runs

[TRTLLM-17033][test] Qwen3.8-Flash-Next accuracy: six cases, each engine running every check that fits

[TRTLLM-17033][test] Qwen3.8-Flash-Next ops: at most three cells per op, one property and one guard test

[TRTLLM-17033][test] Qwen3.8-Flash-Next accuracy: acceptance floors for the fp8 and NVFP4 KV engines

[TRTLLM-17033][fix] PyExecutor admission: the model hook and its tests next to main's KV retention check