Skip to content

[Perf] [GB300] qwen3_5: attention, GDN front and out_proj, shared expert and the MTP draft in the plugin; TP4, and two and four requests at TP8 behind an opt-in - #20

Merged
Oasis-Git merged 28 commits into
sgl-project:mainfrom
YAMY1234:qwen35-gb300-full
Oct 27, 2026
Merged

Oasis-Git merged 28 commits into
sgl-project:mainfrom
YAMY1234:qwen35-gb300-full

Conversation

@YAMY1234

@YAMY1234 YAMY1234 commented Oct 21, 2026 •

Copy link
Copy Markdown
Collaborator

Motivation

This PR adds the Qwen3.5-397B-A17B (NVFP4) MTP megakernel for GB300: the target's verify step and the MTP draft at TP8 across two trays, and TP4 on one tray.

On the SGLang side it needs sgl-project/sglang#45266 (the Qwen3.5 binding, the EAGLE draft entries and the KV-length graph variants). sgl-project/sglang#45474 holds the runtime changes that turn the plugin's GPU savings into speed. The two SGLang PRs are independent; either can merge first.

With both SGLang PRs and this plugin, static decode at 128K context goes from 706.4 / 700.3 to 1291.7 / 1276.0 TPS/User (P50 / P90). That is 101.0% of the development build these kernels were taken from at P50. On InferenceX AgentX (TP8, concurrency 11) the formal score is X 1234.57, against the development build's 1219.51.

Modifications

One commit per optimization. Every slot is opt-in by name in SGLANG_NVIDIA_MEGAKERNEL_SLOTS. "Numeric" slots change the order of floating-point operations; the others are bitwise equal to SGLang's ops.

  1. Foundation: the plugin's Qwen3.5 target spec, a stock step that runs SGLang's ops through the binding, and a parity tool (SGLANG_NVIDIA_MEGAKERNEL_PARITY=1). It adds the first five slots, all bitwise:

    • route: router GEMM and MoE input quantization in one launch;
    • mkm: top-k, FC1, SwiGLU, requantization and FC2 in one launch;
    • gdn_verify: the GatedDeltaNet verify chain;
    • gdn_pdl: its ReplaySSM step as a programmatic dependent;
    • hop: the layer boundaries' all-reduce + residual + RMSNorm on an NVLink multicast hop.
  2. Target slots, one commit each:

    • gdn_push: the GDN out_proj as the hop's push GEMM;
    • gdn_front: GDN projections and conv in one launch;
    • route_shared: the shared expert's gate/up in the router's launch;
    • skv2: full attention on split-KV;
    • mka: the fused attention prologue and epilogue;
    • mka_push: the attention epilogue pushes into the hop.
  3. Draft entries draft_loop / draft_extend on SGLang's ops, then one commit per draft slot:

    • draft_moe;
    • draft_q1: split-KV decode with KV-length graph variants;
    • draft_xchg: fc split, exchange and head;
    • draft_hop: all-reduce points and o_proj push;
    • draft_front: front and FP8 KV write in one launch;
    • draft_skv2.

    A further commit has the draft extend write its token and rows straight into SGLang's outputs.

  4. Refinements, each its own measured optimization:

    • GDN front weights by one bulk copy per CTA;
    • SKV2 and draft q1 launched as programmatic dependents;
    • the final hop releases the draft head early;
    • the draft head yields SMs on hidden ranks;
    • the draft front as a programmatic dependent;
    • the draft o_proj GEMV applies the attention gate.
  5. TP4 on one tray: mkm's engine for intermediate size 256 per rank; the draft hop uses SGLang's o_proj at TP4. gdn_front is TP8-only.

  6. multi_request (opt-in, TP8): slots also cover verify steps of 2 and 4 requests (14 and 28 rows). It is off by default, because it costs batch 1 about 0.3%.

  7. mkm_wide: routed experts of 8–16 rows on a 16-token engine. Requires multi_request.

  8. draft_loop serves two requests under multi_request.

Usage

# TP8 on two GB300 trays, SGLang with sgl-project/sglang#45266 (+ #45474 for its switches)
SGLANG_ENABLE_EAGLE_POST_VERIFY_CUDA_GRAPH=1 SGLANG_ENABLE_GDN_PREFILL_GRAPH_EXTEND=1 \
SGLANG_NVIDIA_MEGAKERNEL_SLOTS=route,mkm,gdn_verify,gdn_pdl,hop,draft_loop,draft_extend,draft_moe,draft_q1,draft_xchg,draft_hop,draft_front,draft_skv2,gdn_push,gdn_front,route_shared,skv2,mka,mka_push \
python -m sglang.launch_server --model-path nvidia/Qwen3.5-397B-A17B-NVFP4-V2 \
  --tp 8 --nnodes 2 --speculative-algorithm NEXTN --speculative-num-steps 6 \
  --speculative-eagle-topk 1 --speculative-num-draft-tokens 7 \
  --attention-backend trtllm_mha --enable-nvidia-megakernel ...

For more than one request at TP8, add multi_request,mkm_wide. At TP4, drop gdn_front.

Accuracy Tests

  • Kernels: the bitwise slots match SGLang's ops bit for bit. The numeric slots match the development build's kernels bit for bit at the served shapes.
    • Kernel tests: 179 passed, 13 skipped (TP4 has no 16-token engine).
    • Parity in the server: the 14- and 28-row steps are 0-diff in every real request's rows.
  • GSM8K (fixed 250-question subset, real acceptance, sampled at temperature 0.6, top-p 0.95, top-k 20):
configuration greedy sampled acceptance length ratio vs reference (greedy / sampled)
TP8, one request: development build (reference) 0.980 0.984 1 / 1
TP8, one request: main + both SGLang PRs + this PR 0.980 (249 / 250 texts identical) 0.976 1.000 / 0.998
TP4 vs the development build's TP4 0.980 vs 0.980 0.984 vs 0.980 1.000 / 1.002
TP8, two requests, multi_request on (vs off) 0.984 vs 0.980 0.980 vs 0.976 1.000 / 0.997

Speed Tests and Profiling

GB300, static decode with forced acceptance 4.80, about 1,780 rounds per server start. TPS/User = 4.80 tokens / round time. "Full stack" is main + sgl-project/sglang#45474 + #45266 + this PR. TP8 rows are the mean of two pairs of trays run in opposite orders.

configuration (TPS/User P50 / P90) 128K 8K 32K
TP8 main 706.4 / 700.3 763.7 / 751.7 729.9 / 722.7
TP8 main + #45266 + this PR 765.7 / 729.6 764.1 / 724.2 766.5 / 721.4
TP8 full stack 1291.7 / 1276.0 1355.0 / 1338.5 1360.4 / 1342.8
TP8 full stack vs the development build, same trays 101.0% (P50) 100.8% / 100.6% 101.9% / 101.6%
TP4 full stack vs the development build's TP4 101.7% / 100.9% 99.4% / 99.7% 100.1% / 100.2%
  • Without #45474, the round becomes host-bound, so this PR's GPU savings mostly do not show.
  • With multi_request on, TP8 at 128K:
    • two requests: 652.4 → 752.8–761.1 TPS/User at P50 (+15–17%), and 783.2–783.5 with the two-request draft_loop (101.4% of the development build);
    • four requests: 521.5 → 559.3–563.4 (+7–8%).

The head also runs on SGLang main after sgl-project/sglang#45112: it reads the renamed shared-expert count and accepts the moved ReplaySSM files. The speed and accuracy runs used SGLang main 28448a86d4 and this PR's code at 3509187d74, an earlier revision. The head adds the multi_request opt-in (off by default, batch-1 tokens unchanged) and the two-request draft_loop, both measured above, and is otherwise the same code.

@YAMY1234 YAMY1234 changed the title [Perf] [GB300] qwen3_5: attention, GDN front and out_proj, shared expert and the MTP draft in the plugin [Perf] [GB300] qwen3_5: attention, GDN front and out_proj, shared expert and the MTP draft in the plugin; TP4, and two and four requests at TP8 behind an opt-in Oct 21, 2026
@YAMY1234
YAMY1234 marked this pull request as ready for review October 21, 2026 17:26
@YAMY1234
YAMY1234 requested a review from Oasis-Git as a code owner October 21, 2026 17:26
…B MTP verify on GB300 TP8

The target-verify step of one request (7 rows) runs the plugin's stock step: SGLang's ops in
SGLang's order over the weights and pools SGLang loaded, read through SGLang's EAGLE binding.
The parity tool (SGLANG_NVIDIA_MEGAKERNEL_PARITY=1) compares it with SGLang's forward bit for
bit and reports how far apart they are. The MTP draft spec covers no step yet, and a slot list
naming verify_sample is refused with the switch that replaces it.

Opt-in slots (SGLANG_NVIDIA_MEGAKERNEL_SLOTS), each bitwise:
- route: the router GEMM and the NVFP4 MoE input in one launch;
- mkm: on route's outputs, the top-10, FC1, SwiGLU, requantization and FC2 in one launch;
- gdn_verify: the GatedDeltaNet verify chain with four launch choices changed;
- gdn_pdl: its ReplaySSM verify launched as a programmatic dependent of the conv;
- hop: the boundary's three fused all-reduce + residual + RMSNorm points on an NVLink
  multicast hop.
…the hop's push GEMM (numeric)

The GatedDeltaNet's static-FP8 out_proj (1-8 rows) runs as one GEMM whose epilogue pushes the
bf16 partial into the decode hop's mailbox, and the MoE input's all-reduce then only consumes it.
The hop's device protocol moves into q35_hop.cuh so other kernels can push from their epilogues.
Slot gdn_push (with gdn_verify and hop); numeric: it sums in another order than FlashInfer's BMM.
… and conv in one launch (numeric)

One launch of 11 clusters runs the static-FP8 in_proj_qkvz and BF16 in_proj_ba projections, the
conv update and an L2 prefetch of the verify's state; the ReplaySSM verify and the gated norm into
out_proj's FP8 input follow as programmatic dependents. Slot gdn_front (with gdn_verify and
gdn_pdl, one request of 7 rows); numeric: the projections sum in another order than SGLang's GEMMs.
…p in the router front's launch (numeric)

The router front's launch also quantizes the rows to the shared expert's static FP8 and computes
its gate/up, so the shared expert on the alternate stream waits for that launch only. Slot
route_shared (with route, up to 8 rows); numeric: the gate/up sums in another order than the FP8
linear's GEMM.
…plit-KV attention (numeric)

After the fused FP8 K / V write, the full attention's core runs SKV2, the split-KV attention of
SGLang branch nvidia/qwen35-agentx-gb300 (a CuTe DSL producer over 148 page splits and a Triton
merge); its KV-length graph variants keep SGLang's attention below 8192 tokens. Slot skv2;
numeric: SKV2 splits the KV and sums in another order than trtllm-gen.
… and epilogue (numeric)

MKA's prologue writes the FP8 query, the bf16 gate and the step's FP8 K / V rows into the pool,
and its epilogue gates, quantizes and runs o_proj into the TP-partial output (the kernels of
SGLang branch nvidia/qwen35-agentx-gb300, built with NVRTC). Slot mka; numeric: its qkv and
o_proj GEMMs sum in another order than SGLang's.
…its rows into the hop

Where the MoE input's all-reduce runs the hop, MKA's epilogue pushes its output fragments into
the hop's mailbox and that all-reduce only consumes them. Slot mka_push (with mka and hop); the
pushed values are mka's own, so it adds no tolerance.
…EAGLE entries on SGLang's own ops

The draft spec offers draft_loop (with begin_draft_round) and draft_extend for one request when
they are requested in SGLANG_NVIDIA_MEGAKERNEL_SLOTS. Each runs inside SGLang's draft graphs over
SGLang's loaded draft modules with SGLang's own ops (bitwise), the base the draft slots replace.
…one launch (numeric)

Inside draft_loop, a one-row draft step's MoE (router GEMV, softmax top-10, routed experts and
gated shared expert) runs as one launch of one CTA per SM on the bf16 weights SGLang loaded.
Draft slot draft_moe; numeric: the dot products sum in another fp32 order, so the draft's claim
turns into TOLERANCE.
…t-KV decode (numeric)

Inside draft_loop, a one-row draft step's decode attention runs SKV2-Q1: trtllm-gen's KV split and
rounding points, the CTAs merged in fp32 by a small Triton kernel. The draft loop's graph variants
keep trtllm-gen below 16384 tokens of KV. Draft slot draft_q1; numeric: only the fp32 association
differs from trtllm-gen.
…y's fc split, exchange and head (numeric)

Inside draft_loop, the greedy one-row draft step boundary runs in two launches instead of seven:
the fc as a row split, the LM head as a GEMV on the vocabulary shard, and one multicast exchange
for the argmax, the positions and the next step's fc. The hidden-half fc partial declares one CTA
per SM, as it launches, so NVRTC's build issues its weight loads together. Draft slot draft_xchg;
numeric: only the fp32 summation order differs from SGLang's fc and LM head.
…points and o_proj push on the decode hop (numeric)

Inside draft_loop, the draft step's two all-reduce + residual + norm points run as hops on the
target's decode hop, in the rounding of the kernels they replace, and o_proj runs as a one-row
bf16 GEMV whose epilogue pushes into the FFN input's hop. Draft slot draft_hop (with hop);
numeric: the GEMV sums in another fp32 order than SGLang's GEMM.
… front writes the FP8 KV in the same launch

After the qkv projection, one Triton launch runs the q / k Gemma RMSNorm, the RoPE and the gate
split, and also writes the FP8 query and the K / V rows into the FP8 KV cache, in place of SGLang's
second launch. Draft slot draft_front; bitwise: the same bytes as SGLang's two launches.
…ention on the split-KV attention (numeric)

Inside draft_extend, the draft extend's decode attention runs on slot skv2's split-KV attention
over the draft's KV pool, with the extend's KV-length graph variants (SGLang's attention below
32768 tokens). Draft slot draft_skv2 (with draft_extend); numeric: SKV2 sums in another order than
trtllm-gen.
…traight into SGLang's outputs

The served draft extend gathers its selected rows into hidden_out, and the exchange writes the
token and its 1.0 into topk_index_out / topk_p_out, so the graph's three closing copies go.
Entry draft_extend with slot draft_xchg; bitwise: the values are the same.
…ghts by one bulk copy into shared memory

Slot gdn_front: each projection CTA brings its 128 KB of packed FP8 in_proj_qkvz weights into
shared memory with one TMA bulk copy issued before its dependency wait, instead of 32 register
loads per thread. The MMA sequence is unchanged, so every output is the same (bitwise).
…tic dependents

Slots skv2 and draft_skv2: the split-KV producer launches as a programmatic dependent of the
attention's front and loads its leading K / V tiles before it waits, and the merge launches as
the producer's dependent. Same arithmetic, same bytes (bitwise).
…rogrammatic dependents

Slot draft_q1: the SKV2-Q1 producer launches as a programmatic dependent and loads its leading
K / V tiles before it waits, and the merge waits before it reads the partials; slot draft_front's
one-row front releases its dependents at entry. The arithmetic is unchanged (bitwise).
…entry, with a shared-memory pad

Slots draft_xchg and draft_hop: the draft step's final all-reduce + norm hop releases the draft
head at its entry, and the head launches with 120 KB of unused shared memory per CTA so that it
starts only on SMs the draft MoE has left. Each CTA's rows and arithmetic are unchanged (bitwise).
…the side-stream fc partial

Slot draft_xchg: on the ranks whose fc slice is in the hidden half, the next step's hidden-half fc
partial launches on 24 CTAs and the draft head on the remaining SMs beside it. Each row is still
one warp's in the same arithmetic, so the logits, the partial and the token are the same (bitwise).
…atic dependent of the qkv GEMM

Slot draft_front: with one row, the front loads its static operands (the norm weights, the M-RoPE
axis map, the cache scales) before its dependency wait and issues every other load right after
it. The arithmetic is unchanged (bitwise).
…utput gate as it stages its input

Slot draft_hop: the one-row o_proj GEMV takes x * sigmoid(gate), computed while it stages x in the
instruction sequence of SGLang's gate kernel, so that kernel no longer runs, and the q1 merge
releases the GEMV at entry. The arithmetic is unchanged (bitwise).
… for intermediate 256 per rank

The spec takes TP sizes 4 and 8. Slot mkm picks its engine from the intermediate size a rank holds,
adding the TP4 engine of SGLang branch nvidia/qwen35-agentx-gb300 (bitwise against trtllm-gen), and
slot draft_hop keeps SGLang's o_proj at TP4. Every other slot builds as at TP8, except gdn_front,
which refuses at TP4.
…s of 2 and 4 requests

With multi_request in the slot list, TP8 also covers the verify steps of 2 and 4 requests (14 and
28 rows), each slot only up to the rows its kernels serve: route 16, hop 15, gdn_push, route_shared
and mka 8, mkm 7, gdn_front and skv2 one request, gdn_verify and gdn_pdl every width. Off by
default; TP4 keeps one request. The parity check counts the requests' own rows apart in graph
steps. Each slot keeps its claim.
…16 rows in one launch

The 16-token engine of SGLang branch nvidia/qwen35-agentx-gb300 (a CTA per (token, top-k) slot, no
clusters) at the TP8 expert shapes runs slot mkm's steps of 8 to 16 rows. Slot mkm_wide (with mkm
and multi_request; vacuous where no covered step is wider than 7 rows); bitwise against
FlashInfer's trtllm-gen MoE.
…ith multi_request

A draft step of two requests runs the split fc, the hop's all-reduce points and the two-token MoE;
q1, the o_proj push and the exchange stay one-request. It serves them where multi_request covers
the verify steps of several requests (TP8); TP4 keeps one request. Numeric as the draft slots.
# Conflicts:
#	.github/CODEOWNERS
#	README.md
#	sgl_nvidia_mega/plugin.py
#	test/unit/test_contract.py
#	test/unit/test_plugin.py
@Oasis-Git
Oasis-Git merged commit 9a21cb0 into sgl-project:main Oct 27, 2026
2 checks passed
nvrohanv added a commit to nvrohanv/sgl-nvidia-mega that referenced this pull request Oct 29, 2026
Main's SGLANG_MIRRORS check (sgl-project#20) digests a
function's tokens in another encoding than this branch's copy, which the
merge replaced with main's. Pin the 18 GLM-5.3 mirrors with main's
mirror_digest of the same SGLang code: no mirrored body changes, only how
its digest is written.
YAMY1234 added a commit that referenced this pull request Nov 10, 2026
… NVFP4 MTP target verify and NextN draft (#24)

* [Feat] [B300] glm5_3: NVIDIA megakernel specs for GLM-5.3 NVFP4 MTP target verify and NextN draft on B300 TP8

Add sgl_nvidia_mega/glm5_3: the GLM-5.3 target spec for SGLang's megakernel
layer (MTP verify steps of 1 to 4 requests, 4 to 16 rows, on 8 x B300 at
TP 8), its step functions with stock paths adapted from SGLang's model code,
and the CuTe DSL attention, small-M NVFP4 MoE and all-reduce hop kernels.
With --enable-nvidia-megakernel every target slot runs (attention, moe,
hops, qkv_a_small_m_shared, dnorm); SGLANG_NVIDIA_MEGAKERNEL_SLOTS is a test
and bisect override. Needs SGLang's DSA binding view.

Add the GlmMoeDsaForCausalLMNextN draft spec (slot draft, on by default) and
its CuTe DSL kernels: the embedding gather and eh norms in one kernel, the
column-sharded eh projection gathered by the target's all-reduce hop, the
distributed top-1, and the draft layer's MLA prologue, in every captured
draft graph. Needs embedding replication (with the flag, SGLang's megakernel
layer sets SGLANG_ENABLE_EMBED_REPLICATION when it is unset), greedy draft
acceptance and SGLang's EAGLE draft graph capture check; the contract test
skips the check's marker, naming it, on an SGLang without it.

At startup rank 0 logs the covered verify widths (1-4 requests) and the
derived weight copies' memory per GPU, and warns when
--max-running-requests is set above 8, where the megakernel helps little.

Also the README section, a NOTICE paragraph, a CODEOWNERS line and the
contract test's checks of the DSA binding view reads and the mirrored SGLang
code; the GLM-5.3 tests are filed separately.

The same specs serve GLM-5.2 NVFP4, which has GLM-5.3's architecture
(GlmMoeDsaForCausalLM) and the same value of every config field the target
and draft specs pin. The spec names and the plugin's log, refusal and rank
agreement texts name both models, and the module path stays glm5_3.
test/unit/test_glm5_3_glm5_2_config.py runs the published config of
nvidia/GLM-5.2-NVFP4 (revision 53e0691e21895a3863a606dfd12910c69eba94ab)
through SGLang's requirement check for both specs, and checks that a config
that differs in any pinned value, or lacks one, is still refused.

The package imports SGLang's megakernel interface from its location since
sgl-project/sglang#45389 (sglang.kernels.plugin_kernels.nvidia,
sglang.srt.kernel_plugins.nvidia) and falls back to sglang.srt.megakernels
on an older SGLang, as deepseek_v41 does.

* docs: correct GLM GPU test availability

* [Feat] [B300] glm5_3: serve MTP with 5 draft steps (6 verify rows a request)

The GLM-5.3 / GLM-5.2 specs served MTP / EAGLE with 3 draft steps only, so
the 5 / 1 / 6 setup of SGLang's GLM-5.3 and GLM-5.2 B300 NVFP4 low-latency
recipes stopped at startup with --enable-nvidia-megakernel.

The verify widths, the MoE row map and the draft eh rows are now derived
from the verify rows a request T (speculative_num_draft_tokens =
speculative_num_steps + 1, top-1 drafts), and the specs serve the tested
values 4 and 6 (plan.SERVED_TOKENS_PER_REQUEST); any other value is refused
at startup, naming them.

- Widths: every whole number of requests up to the kernels' 16 rows
  (plan.supported_rows): 4 / 8 / 12 / 16 rows at T 4 as before, 6 / 12 rows
  (1-2 requests) at T 6. Steps of more requests run SGLang's path.
- MoE: the engine takes at most 4 row groups of at most 4 rows, token sets
  over one expert list rather than requests. moe_split groups the rows by 4
  with the remainder last: T 4 keeps (4,) to (4, 4, 4, 4); 6 rows run on the
  two-group engine as (4, 2) and 12 rows on the four-group engine as
  (4, 4, 4). The one-group engine holds 4 rows, so it cannot take 6.
- Draft: the eh_proj shard serves the draft decode's 1-4 rows and the draft
  extend at the covered widths, so T 4 keeps its rows and T 6 adds 6;
  remove the unused max_requests.
- Attention, the hops and the draft attention already take any row count
  up to 16 on their 8- and 16-row forms; no kernel changes.

At T 4 every width, row map, eh row set and kernel form is unchanged.

* [Perf] [B300] glm5_3: the one-request fast forms at 6 verify rows

With 5 draft steps a one-request verify step has 6 rows, past the 4-row
limit of the forms that serve one request of 3 draft steps. Serve them at up
to 6 rows:

- the split MoE fold hop (SPLIT_MAX_ROWS 6; still taken only where the
  cooperative occupancy holds 2 x rows + 1 clusters of 8 CTAs, probed at
  build);
- the plain all-reduce hops on the dense layers (DENSE_MAX_ROWS 6);
- the 16-warp sparse MLA (one CTA per SM; its kernel already takes any row
  count and gives the 8-warp form's bytes).

No width of 4 verify rows a request has 5 or 6 rows, so those steps keep
their forms. The kernels are unchanged.

* docs: glm5_3 MoE groups and eh_proj rows with 6 draft tokens

The README's GLM-5.3 slot table named the MoE engine forms by requests
and listed the eh_proj rows of 4 draft tokens only. Name the forms by row
groups (6 rows run as 4 + 2, 12 as 4 + 4 + 4) and add the eh_proj rows of
6 draft tokens (6 and 12).

* [Fix] glm5_3: pin the SGLang mirrors with main's mirror digest

Main's SGLANG_MIRRORS check (#20) digests a
function's tokens in another encoding than this branch's copy, which the
merge replaced with main's. Pin the 18 GLM-5.3 mirrors with main's
mirror_digest of the same SGLang code: no mirrored body changes, only how
its digest is written.

* [Fix] glm5_3: re-pin four SGLang mirrors for current SGLang main

Four mirrored SGLang bodies changed on SGLang main. Each pin is now a
tuple of the digests before and after the change, so the check passes on
either SGLang:

- DeepseekMLAForwardMixin.forward_absorb_prepare (sgl-project/sglang#46359):
  every changed line is on the --dcp-replicate-q-proj path, which runs only
  with decode context parallelism; the specs refuse it, and the stock
  attention step does not copy those lines.
- DeepseekSparseAttnBackend._forward_trtllm (sgl-project/sglang#46990): the
  KV store and read go through the pool's layer router; on the DSA pool, a
  plain pool, that is the same set_mla_kv_buffer and get_key_buffer call
  with the same arguments, which the stock step makes.
- DeepseekSparseAttnBackend.forward_decode: sgl-project/sglang#46990 as
  above, and sgl-project/sglang#44322, a triton branch for ROCm with
  decode context parallelism.
- LogitsProcessor.forward (sgl-project/sglang#43449): the new lines run only
  with --enable-return-aux-hidden-states, which turns speculative decoding
  off, so never with the NextN draft; the output's new last_hidden_states
  stays None.

The copied steps stay as they are. The other 14 mirrored bodies are
unchanged on SGLang main.

* [Fix] glm5_3: adapt the specs to SGLang's megakernel API version 2

SGLang main's megakernel contract is version 2: a built object may also
offer the EAGLE entries draft_loop, draft_extend and begin_draft_round and
attention graph_variants, which SGLang reads from every built object. The
GLM-5.3 / GLM-5.2 target and draft built objects offer none of them (None:
SGLang's own path, as before), and the specs declare API version 2, as
main's DeepSeek-V4.1 and Qwen3.5 specs do.

* docs: glm5_3 boundary docstring says plain dense hops up to 6 rows

DENSE_MAX_ROWS is 6 since the 5-draft-step change; the module docstring
still said 4.

* test: glm5_3 pins the forward_mla mirrors after sglang#45946

sgl-project/sglang#45946 adds SAIL-only fp8 conditions to
forward_absorb_prepare / forward_absorb_core and a FlashMLA-backend-only
llama_4_scaling step to forward_absorb_prepare; the DSA path the spec
serves is unchanged. Pin the new digests next to the old ones.

* [Feat] [Rubin] glm5_3: serve SM107, TP8 across two trays, and nodes without NVLink multicast

Port the Rubin pieces of #19 (glm_dsa) to the
GLM-5.3 / GLM-5.2 specs. The kernels are the same; only the admission changes.

- spec, draft: admit sm_107 (Rubin / VR NVL72) next to sm_103. The draft
  takes the target's ISAs.
- spec, draft: declare Feature.MULTI_NODE_TP, so TP8 can span the two 4-GPU
  trays of an NVL72 rack (GB300 NVL72, Rubin VR NVL72), as the Qwen3.5 and
  Kimi K3 specs do.
- runtime: SGLANG_NVIDIA_MEGAKERNEL_GLM_ALLOW_NO_AR_FUSION=1 (off by default)
  lets the model check and the per-step coverage pass where SGLang's FlashInfer
  all-reduce + RMSNorm fusion does not apply (no NVLink multicast). The stock
  residual steps then run SGLang's unfused all-reduce, which is what its
  boundaries run there, and slot hops runs its own symmetric-memory all-reduce.
  Without the variable nothing changes.
- README: the GPUs and the two-tray launch. test/unit/test_glm5_3_rubin.py
  covers the three changes on CPU.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: Run the NextN draft-extend MLP on the select_index rows only

The draft extend runs the NextN layer on every verify row of each request, but
the next draft step reads only the select_index row (one per request); the
logits processor already prunes to those rows before the LM head. After the
post-attention norm the draft step now keeps those rows of the MLP input and of
the residual SGLang's prepare wrote, so the MoE, its all-reduce and the final
norm run on B rows instead of B * tokens (6 -> 1 and 12 -> 2 rows at EAGLE
5/1/6). The attention still writes every row's KV and indexer key. The hidden
states handed to the logits then hold only the selected rows, with a
LogitsMetadata that does not select again.

- Slot draft_rowselect (shipped, a draft slot, needs slot draft).
- Gate (draft_step.select_index_rows): DRAFT_EXTEND_V2, no logprobs, hidden
  capture not FULL, no padded-row mask, a 1-D select_index with 0 < B < N
  rows; the NextN MLP is DeepseekV2MoE without an all-to-all. Host-only facts,
  fixed per captured shape. draft_attention.select_rows raises on any other
  stream layout.
- draft_attention.decoder_layer_forward takes the rows to keep (ffn_rows) and,
  at a width the draft attention family does not serve, SGLang's own attention
  call.
- Rank 0 logs once per captured draft graph its attention and MLP rows.

Numerics: the selection itself is exact (the kept rows equal the full-row
stream's rows bit for bit through SGLang's residual stream and final add +
RMSNorm). The draft MoE and its router pick their kernels by the row count, so
the MLP output on B rows is not bit-identical to the same rows of the N-row run
(draft.rowselect.mlp.* in the draft NUMERICS). Draft-only: target verification
decides every emitted token.

Tests: test/unit/test_glm5_3_draft_rowselect.py (CPU; the logits hand-off
needs CUDA), test/kernels/test_glm5_3_draft_rowselect.py (4 GPUs).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Feat] glm5_3: TP4 profiles, verify steps composed per width from kernel units

The spec serves TP8 at 4 or 6 tokens per request up to 16 rows. It now also
accepts TP4 at 3 to 6 tokens per request up to 48 rows (one GB300 tray: the
InferenceX AgentX TP4 aggregate and PD decode-role configurations), and serves
those widths with kernel units composed per step; the TP8 path
(plan.build_providers, the executor, the families) is unchanged.

- profile.py: the facts a build is decided on (TP, heads per rank, tokens per
  request, widths, device, features) and the accepted table: the TP8 envelope
  (plan.py) plus TP4 at 3-6 tokens, 48 rows. Profile.units says which build
  runs.
- registry.py: candidates claim steps of a group atomically, with coverage per
  (profile, rows, mode) and contract tokens; per width the first preferred,
  switched-on, covering unit takes each step, the rest run the stock body. A
  decline (coverage) runs stock and is logged once per width; a family that
  covers a width but fails to build raises on every rank. A mixed group is a
  subclass of the stock group with the claimed steps bound to their units, so
  the stock step bodies and Providers.describe see them.
- units.py: the TP4 preference table. Every unit is a slot (runtime.LL_SLOTS,
  shipped; OPT_IN_SLOTS only when listed), selected by
  SGLANG_NVIDIA_MEGAKERNEL_SLOTS as the TP8 families are; "" is parity (every
  accepted width on the stock bodies).
- runtime/composed.py: the executor (GlmMegakernelVerify: its layer loop,
  first-capture checks and parity switch) on the composed step groups; its
  logits take ll.gemm's lm_head role and share the model processor's gatherer.
  runtime.build hands a TP4 profile to it.
- spec.py: tp_sizes {4, 8}, the (TP, tokens) pairing checked against the
  profile table (profile_reasons), max_rows 48, the units' named tolerances.
- ll/__init__.py, ll/jit.py: the units' package (optional modules, the
  ll.linear GEMM entry point, named tolerances) and the JIT loader for its .cuh
  kernels (SGLang's load_jit; refuses to compile under capture; failures
  raise).

Tests: test/unit/test_glm5_3_{profile,registry,spec,ll_jit,verify_logits}.py.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: TP4 small-M BF16 GEMMs on PDL split-K and TGV (ll.gemm)

The TP4 roles' small-M BF16 GEMMs that SGLang leaves on cuBLAS launch without
programmatic dependent launch. ll.gemm moves them, at the row counts where a
GB300 crossover bench measured a win in the GEMM's model context, to a PDL
kernel (slot ll.gemm.<role> each):

- router (FP32 logits): split-K at 17-96 rows (the tiny GEMM keeps <= 16);
- indexer wk|weights_proj: split-K at 1-64 rows;
- o_proj TP4: TGV 1-72; dense_down TP4: TGV 1-48; NextN eh_proj: TGV;
- lm_head TP4: TGV 1-60; measured but deferred (run only when listed): the
  TP8 o_proj / q_b shapes.

gemm_plan.py holds the plan and the named tolerances. The plan applies only on
SM103 with 152 SMs; elsewhere, and for quantized or biased weights or unplanned
rows, every role runs SGLang's own GEMM for the module, bit for bit.

Split-K is a vendored copy of FlashInfer v0.7.2's cluster split-K kernel with
FP32 output / any M and synchronization fixes (owner wait at .acquire.cluster,
phase-0 wait before the TMEM release, tcgen05 thread-sync fences; no arithmetic
change). FlashInfer's own kernel is used only when it carries those fixes
(SplitKTactic.SYNC_FIXED); SGLANG_NVIDIA_MEGAKERNEL_GLM_GEMM_SPLITK pins
auto / flashinfer / vendored. TGV is SGLang's cutedsl_bf16_gemm.

Tests: test/unit/test_glm5_3_ll_gemm_plan.py (CPU: tactics, engagement only on
SM103 / 152 SMs, slots, FlashInfer detection, stock dispatch);
test/kernels/test_glm5_3_ll_gemm.py (1 GPU: path per role and row, tolerances
vs cuBLAS and FP64, graph replay, PDL edges, launch stress).
benchmark/bench_glm5_3_ll_gemm.py re-checks the crossover.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: TP4 sparse MLA flash decode as unit ll.sparse_mla_fd

The sparse_mla and uv_proj steps of the TP4 verify attention and of the NextN
draft layer's attention on a flash-decoding CUDA kernel over the top-k gathered
FP8 KV rows, instead of TRTLLM-Gen's sparse MLA decode + torch.bmm. The step
reads the running forward's ForwardView (view.dsa, view.kv_pool,
view.metadata), as the stock step does.

- Kernels (ll/csrc/attention/glm_sparse_mla_fd_*.cuh, adapted from
  TensorRT-LLM's sparse MLA flash decode): one 16-CTA cluster per row, 8 or 16
  heads; a v_b fold instance (8 heads, bitwise cuBLAS's BMM); a multi-tile mode
  (clusters of 2-8 CTAs, 8 heads).
- Launch plan (GB300, 152 SMs): one-tile up to 7 rows; built at compile; the
  co-resident cluster counts are agreed over the TP group (MIN); a build
  failure on any rank raises on every rank.
- Covers sm_103 with 152 SMs, 16 heads (TP4), <= 7 rows; declines with
  TRTLLM-Gen skip-softmax decode on.
- Tolerance ll.sparse_mla_fd.max_abs = 0.0625 against TRTLLM-Gen (measured
  0.052 / 0.055 at 8 / 16 heads).

16 heads: -2.1 to -2.8 us per layer at 5 / 6 rows
(benchmark/bench_glm5_3_sparse_mla.py).

Tests: numerics vs TRTLLM-Gen and an FP64 oracle, the provider steps vs the
stock steps, CUDA-graph replay and the PDL chain (1 GPU); the launch plan and
rank agreement (CPU).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: TP4 attention front on low-latency fused-A and a one-launch MLA KV half

The unit ll.attention (groups attention and draft_attention) serves the
attention front of a TP4 rank (16 heads, SM103, 1-48 rows, every step mode) by
claiming qkv_a / prologue / index_k / rope_cache of AttentionStep:

- qkv_a: SGLang's min-latency fused-A kernel with a larger stage ring and a
  pipelined MMA loop from 9 rows, up to 16 rows (bit for bit SGLang's kernel); a
  split-K form at 17-24 rows when one wave fits (ll.attention.qkv_a_wide.max_rel
  of fp32); stock above 24 rows.
- prologue: stock q_a norm; q_b on the pipelined form at 9-16 rows (bitwise);
  the kv_a norm deferred into rope_cache when folded.
- index_k: the stock body with its key / head-gate projection through
  ll.linear("wk_weights_proj").
- rope_cache: one PDL launch for the kv_a norm (folded), RoPE and FP8 of q and
  k and the MLA cache append, bit for bit FlashInfer's norm and
  mla_rope_quantize_fp8 and SGLang's set_mla_kv_buffer (the cache through the
  step's ForwardView).

The fold (slot ll.attention.kv_norm_fold) engages only when FlashInfer's norm
backend, read from its public surface, is CuTe DSL and a probe at build gives
the stock kv_a norm's bits; the family logs which way it went. fused_a_ll.cuh is
adapted from SGLang's dsv3_fused_a_gemm.cuh.

Tests (1 GPU): fused-A forms bitwise at 1-16 rows and within tolerance at
17-24, graph replay; the KV half bitwise (q and every cache byte); the stock
AttentionStep vs ll.attention on one TP4 layer at 5-48 rows, eager and
replayed. CPU: the norm-backend decision, coverage and composition.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: TP4 all-reduce + residual + RMSNorm hops on a plugin-owned Lamport buffer

The unit ll.hops serves the three hop steps (HopSteps) of the TP4 verify step:

- csrc/distributed/glm_ll_mnnvl_oneshot.cuh: one-shot / two-shot MNNVL
  all-reduce + residual + RMSNorm with a MoE-finalize prologue and an NVFP4
  epilogue (builds with nvcc 13.0 and 13.4). Every wait on a peer is bounded:
  past the deadline it writes a host-mapped error record and traps (fail-stop).
- ll/transport.py: the Lamport buffer and flags are the plugin's own (torch
  symmetric memory, or FlashInfer's McastGPUBuffer) in FlashInfer's MNNVL
  layout; no FlashInfer workspace or flags are shared. Modules load through
  ll.jit; a build failure raises on every rank.
- ll/hops.py: the GB300 launch plan per world size and row count (W=4: one-shot
  up to 36 rows, two-shot 37-48), the family (deadline min(20 s, the build
  context's), o_proj through ll.linear) and its step group. Covers the TP4
  verify step on sm_103 up to 48 rows; provides hops:finalize and
  mlp_input:nvfp4 (written only when a taken MLP unit consumes it).

Tests: the plan tables (CPU); at W=4 every row 1-48 bit-identical to
FlashInfer's one-shot / two-shot, the finalize hop bit-identical to
moe_finalize_fuse_shared + the hop, a protocol mix eager and in a CUDA graph;
fail-stop of every cross-GPU wait; a two-tray multicast probe.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: TP4 MoE and dense MLP steps with the deferred finalize (ll.moe)

The units ll.moe.nvfp4 and ll.moe serve the MLP step (MlpSteps) of the TP4
verify step and leave the MoE finalize to the next hop:

- MoE layers: the router GEMM (ll.linear("router"), FP32 logits), then
  FlashInfer's public trtllm_fp4_block_scale_moe with do_finalize=False and the
  layer's processed NVFP4 weights (SGLang's
  fused_experts_none_to_flashinfer_trtllm_fp4 call for this configuration), and
  the shared expert on the layer's alternate stream; returns a deferred
  finalize for the next hop.
- ll.moe.nvfp4 takes the routed experts' FC1 input from the post-attention
  hop's NVFP4 epilogue; ll.moe quantizes its own input.
- Dense layers: the stock gate_up_proj and SiLU-and-mul, then
  ll.linear("dense_down"), un-reduced.
- Opt-in slot ll.moe.alt_stream_priority: the shared expert on a stream of the
  greatest priority at <= 6 rows. It gains only with FlashInfer's small-batch
  routing kernel (TP4 5/1/6: -1.82% step at 6 rows); on stock FlashInfer 0.7.2
  it loses (+1.2% at 5 rows, TP4 4/1/5, n = 2; +2.4% at 6 rows), so it runs only
  when listed in SGLANG_NVIDIA_MEGAKERNEL_SLOTS.
- Both require hops:finalize. A layer that does not fit declines the family
  with the layer and the reason.

Tests (4 GPUs): the copied call equals SGLang's runner bit for bit at 5-48 rows;
the layer (hop + MoE + finalize hop) equals the stock layer bit for bit on the
hop's FP4 input inside FlashInfer's one-shot band; CUDA-graph replays equal
eager; the priority stream keeps the bits.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Perf] [GB300] glm5_3: TP4 NextN draft composed from units (P6a eh_proj, LL attention, row select)

The draft step (draft_step.FusedDraft) serves TP8. At a TP4 profile the draft
spec now builds draft_units.ComposedBuiltDraft: each draft graph (draft decode
at one row per request, draft extend at the verify rows) is composed through
the registry from the draft step groups (draft_front, draft_eh,
draft_attention, draft_mlp, draft_top1):

- draft.front (slot draft): the draft step's fused embedding + eh norms (the
  kernel does not depend on the TP size);
- ll.draft_eh (P6a, ll/draft.py): eh_proj row-parallel, each rank's K window
  times its weight columns. With the target's LL hops (found through the
  shared-table registry, runtime.register_target_hops) the partials are summed
  and the draft layer's input norm applied in one hop launch, and the layer
  starts from that (draft_attention.PreparedInput: the attention prepare is
  skipped); otherwise SGLang's all-reduce and the stock norm. Per draft forward
  at TP4, L2 flushed: 20-26 us vs 39-41 us for the replicated GEMM;
- the draft layer's attention on ll.attention and ll.sparse_mla_fd (the stock
  DraftAttention on the step's ForwardView elsewhere);
- draft_rowselect in the draft extend (draft_step.select_index_rows);
- the LM head on ll.gemm's lm_head role; the top-1 stays SGLang's (at TP4 its
  one-tray multimem gather is as fast as the distributed argmax or faster).

The composed draft is built when the target shares its embedding and LM head
(set_embed_and_head, before the KV pool is sized), so the LM-head role is built
on the target's head. Draft graphs outside the profile's widths run SGLang's
draft forward. The TP8 draft keeps the TP8 widths (<= 16 rows).

Tests: test/unit/test_glm5_3_draft_units.py and test_glm5_3_draft_eh_prepared.py
(CPU: widths, compositions, slots, the hand-off to the logits processor),
test/kernels/test_glm5_3_draft_eh_proj.py (4 GPUs: numerics at 1-48 rows, the
hop sum + norm, graph replay).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Test] glm5_3: Serving tests and decode A/B configuration at the AgentX TP4 / TP8 points

- test/e2e/glm5_3_server.py, test_glm5_3_server.py: servers at the InferenceX
  AgentX configurations (tp4-416: aggregated TP4, EAGLE 4/1/5, HiCache;
  tp8-516: aggregated TP8 on two trays, EAGLE 5/1/6, HiCache; tp4-516: the
  disaggregated decode role, TP4, EAGLE 5/1/6, standalone): on / off (the
  active line, covered decode graphs, GSM8K-200 within 3 points), parity
  (SGLANG_NVIDIA_MEGAKERNEL_SLOTS="" vs flag off, greedy ids and logprobs
  bitwise) and HiCache load-back (load-back output equals the device-hit
  output).
- benchmark/README.md: the GLM decode A/B protocol at those configurations.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Feat] [GB300] glm5_3: Serve the TP4 decode role of a PD-disaggregated deployment

A PD deployment of GLM-5.2 on GB300 (the InferenceX AgentX 1p5d-c30 / 1p3d-c40
points) runs a TP1 PP4 prefill role and TP4 EAGLE 5/1/6 decode roles. The spec now
lists Feature.PD_DISAGGREGATION, so a decode role serves its verify steps on the
TP4 kernel units: the KV arrives through the transfer backend (NIXL) into pool
slots before a request's first verify step, outside the covered step, and the
decode radix cache is not a framework feature. The TP8 envelope keeps PD refused
(pd_reasons). The prefill role runs without the flag (TP1 and PP are refused
anyway).

Tests: test/unit/test_glm5_3_spec.py (the decode role accepted at TP4, refused at
TP8); test/e2e/test_glm5_3_server.py TestGlm53PDDecode (GSM8K through a running
1P1D router, decode engaged, prefill not, no NIXL transfer failure).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Test] glm5_3: Remove the tests this PR added (repo policy: no tests in sgl-nvidia-mega)

Per review, sgl-nvidia-mega carries no tests. Remove the 28 test files this
PR added under test/e2e, test/kernels and test/unit (the GLM-5.3 TP4, draft
row select and PD decode tests and their helpers glm5_3_server.py and
glm5_3_slots.py), and drop the benchmark README's pointer to the removed
test/e2e/glm5_3_server.py. Tests that already exist on the base branch are
left untouched.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* glm5_3: Keep SGLang's copyright notice in the headers of files adapted from SGLang

Apache-2.0 requires derivative files to retain the copyright notices of the
work they adapt. Following the OSRB guidance, add
"Portions Copyright 2023-2024 SGLang Team (Apache-2.0)" after the NVIDIA SPDX
lines of the glm5_3 files adapted from SGLang, and name the origin of
ll/moe.py, whose routed-expert call copies SGLang's. fused_a_ll.cuh copies
most of SGLang's dsv3_fused_a_gemm.cuh, so it keeps that file's header
unchanged and adds the NVIDIA and SGLang lines below it. Header lines only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Fix] glm5_3: pin the SGLang mirrors for main's SGLang (cf6133a)

main's unit tests check out SGLang cf6133a, where eight glm5_3 mirrors no
longer matched. Each change keeps the covered steps as mirrored, so the new
digest is added beside the old ones (both SGLangs pass):

- forward_absorb_prepare / forward_absorb_core (#44503, #38943): the indexer,
  the unfused RoPE and the W_UV projection moved into the run_mla_indexer,
  prepare_mla_qkv and project_mla_values hooks, whose defaults are the old
  calls (no SGLang class overrides them); the DCP extend gather is out of the
  spec's setups.
- Indexer.forward_cuda, DeepseekSparseAttnBackend._forward_trtllm (#46890):
  prefill-CP only.
- DeepseekV2DecoderLayer.forward (#46857), in the hops, the MoE step and the
  draft attention: an assert under CP decode attention TP, which the spec
  refuses.
- LogitsProcessor.forward (#47200): vocab-sharded logits for a target verify
  only; the draft steps the mirror covers take the gather as before.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

* [Doc] glm5_3: Document the TP4 split-K GEMM source switch in the README

SGLANG_NVIDIA_MEGAKERNEL_GLM_GEMM_SPLITK picks the split-K source of the TP4
small-M GEMMs (auto, flashinfer or vendored); document it next to the TP4
units, as the README documents Qwen3.5's SGLANG_NVIDIA_MEGAKERNEL_MOE_TACTICS.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

---------

Co-authored-by: Yangmin Li <yangminl@nvidia.com>
Co-authored-by: shuw <shuw@nvidia.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kaixi Matteo Chen <kaiximatteoc@nvidia.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants