Skip to content

[Feat] [NVIDIA] glm5_3: NVIDIA megakernel specs for GLM-5.3 / GLM-5.2 NVFP4 MTP target verify and NextN draft - #24

Merged
YAMY1234 merged 32 commits into
sgl-project:mainfrom
nvrohanv:glm5_3-megakernel-specs
Nov 10, 2026
Merged

YAMY1234 merged 32 commits into
sgl-project:mainfrom
nvrohanv:glm5_3-megakernel-specs

Conversation

@nvrohanv

@nvrohanv nvrohanv commented Oct 24, 2026 •

Copy link
Copy Markdown
Contributor

Motivation

At one to four concurrent requests, a GLM-5.3 NVFP4 MTP decode step on 8 x B300 is bound by launch gaps and small-M GEMMs. Each verify step runs 78 layers of attention, MoE and TP all-reduces at 4 to 16 rows. The NextN draft step runs SGLang's forward, with a logits all-gather over the full vocabulary and separate launches for the embedding, the eh norms and the eh projection.

This PR adds two specs through SGLang's megakernel layer (--enable-nvidia-megakernel, sgl-project/sglang#44945). They follow the layout of the DeepSeek-V4.1 spec:

  • Target spec (GlmMoeDsaForCausalLM): its executor runs CuTe DSL kernels at the covered verify widths. Each kernel has a stock path that reproduces SGLang's verify step bit for bit.
  • Draft spec (GlmMoeDsaForCausalLMNextN): runs the draft step inside the draft-decode and draft-extend CUDA graphs.

Both specs serve MTP with 3 draft steps (4 draft tokens) and with 5 draft steps (6 draft tokens, 5/1/6), the draft settings of SGLang's low-latency GLM-5.3 and GLM-5.2 NVFP4 recipes for B300. GLM-5.2 has GLM-5.3's architecture and the same value for every config field the specs pin, so the same specs serve GLM-5.2-NVFP4.

The same specs also admit:

  • GB300 at TP 4 (one tray);
  • TP 8 across the two trays of one NVLink domain (GB300 NVL72, Rubin VR NVL72);
  • Rubin (SM107), which runs the same kernels.

The accuracy and speed results below are on 8 x B300 at TP 8. At one user on an agentic coding trace, this gives 1.43x stock p90 interactivity (Speed Tests below).

Depends on:

Launch: the README's GLM-5.3 / GLM-5.2 command, which is the checkpoint's serving flags plus --enable-nvidia-megakernel. It needs no slot list and no environment variable.

On two trays, launch the same command on each tray with --nnodes 2 --node-rank <0|1> --dist-init-addr <tray-0 host>:<port>. The covered steps need SGLang's FlashInfer all-reduce + RMSNorm fusion, which needs NVLink multicast. Where multicast is not available, SGLANG_NVIDIA_MEGAKERNEL_GLM_ALLOW_NO_AR_FUSION=1 runs the covered steps on SGLang's unfused all-reduce and the plugin's own symmetric-memory hops.

Modifications

Target spec (sgl_nvidia_mega/glm5_3/)

  • spec.py checks the served setup at argument resolution, before any weight load:

    • SM103 or SM107; BF16; full decode CUDA graphs;
    • TP 8, on one node or across two trays of one NVLink domain, or TP 4 on one GB300 tray. TP 4 is SM103 only: on SM107, a TP 4 build serves no width and stops, naming the reason;
    • EAGLE / NEXTN with topk 1: 4 or 6 draft tokens at TP 8, and 2 to 5 draft steps at TP 4;
    • FP8 E4M3 KV cache with page size 64;
    • the DSA backend with the TRT-LLM decode kernel;
    • the FlashInfer TRT-LLM MoE runner with shared-experts fusion off;
    • the checkpoint shapes and routing pinned in HF_CONFIG;
    • the validated CuTe DSL version.

    Any other setup stops at startup and names each reason; there is no silent fallback. NUMERICS holds one tolerance per kernel family.

  • plan.py, runtime/: the build and the executor.

    • Build: allocates the buffers and derived weights and compiles every kernel at load, before the KV cache is sized.
    • Covered steps: a target-verify forward captured in a full decode graph.
      • At TP 8: 1–4 requests (4 / 8 / 12 / 16 rows) at 4 draft tokens, or 1–2 requests (6 / 12 rows) at 6 draft tokens.
      • At TP 4: up to 48 rows (1–8 requests at 6 rows a request).
    • Everything else runs SGLang's forward.
    • First capture of each width: every rank checks the facts the kernels need (for example, an interleaved RoPE). A failure on any rank stops every rank.
  • sm103/: the CuTe DSL kernels.

    • Attention: the small-M q_a / kv_a projection, MLA prologue, indexer query, sparse MLA decode and value up-projection.
    • MoE: a small-M NVFP4 engine for the routed experts and the BF16 shared expert, in groups of up to 4 rows (6 rows run as 4 + 2, 12 rows as 4 + 4 + 4).
    • Hops: all-reduce + residual + RMSNorm hops over symmetric memory. They cover o_proj with the router on MoE layers, the MoE finalize folded into the MLP-output all-reduce, and the final norm.
  • ll/: the TP 4 kernels and the TP 8 draft-decode pieces. They cover sparse MLA flash decode, the TP 4 attention front, registered small-M GEMMs, the TP 4 hops and MoE, and the draft's pieces. Their CUDA sources are under ll/csrc/. fused_a_ll.cuh is adapted from SGLang's fused-A GEMM and keeps its TensorRT-LLM origin header.

  • Other serving modes: HiCache (--enable-hierarchical-cache) is supported, because the covered steps read only the running requests' device KV pages. At TP 4, so is the decode role of a PD-disaggregated deployment.

  • attention/, moe/, boundary/, logits/: the step functions and their stock bodies, adapted from SGLang's model code. Each adapted file names its origin, and NOTICE has one paragraph for them.

  • Memory and startup:

    • The derived weight copies take about 20 GiB per GPU, of which the routed FC2 records are 15.8 GiB.
    • Rank 0 logs the copies' size and the covered widths.
    • Rank 0 warns when --max-running-requests is above 8: larger verify steps run SGLang's path while the copies still reduce KV-cache memory.
  • Cross-GPU wait: the hops' wait bound is compiled in at 20 s, so the spec refuses a --watchdog-timeout below 40 s (the default of 300 passes).

Slots

--enable-nvidia-megakernel runs every slot:

  • attention, moe and hops;
  • qkv_a_small_m_shared: the small-M q_a / kv_a projection on layers that reuse the previous layer's top-k;
  • dnorm: the deferred input norm;
  • draft;
  • draft_rowselect (needs draft): in each draft-extend graph, the draft layer's MLP, its all-reduce and the final norm run only on the rows the draft's LM head reads, one per request. The kept rows match the full-row path bit for bit.

SGLANG_NVIDIA_MEGAKERNEL_SLOTS overrides the list, for tests and bisecting only. A list that is missing a dependency stops the server at startup and names it.

Other settings, none of them needed for serving:

  • SGLANG_NVIDIA_MEGAKERNEL_GLM_ALLOW_NO_AR_FUSION: the opt-in for two trays without multicast (Launch, above).
  • SGLANG_NVIDIA_MEGAKERNEL_PARITY: the test-only parity mode used in Accuracy Tests.
  • SGLANG_NVIDIA_MEGAKERNEL_GLM_GEMM_SPLITK (auto / flashinfer / vendored): selects the TP 4 split-K GEMM source.
  • SGLANG_NVIDIA_MEGAKERNEL_GLM_FORCE_BUILD_ERROR: a test hook that fails a named build on a named rank.

Draft spec (draft.py, draft_step.py, draft_attention.py, sm103/draft/)

In every captured draft graph:

  • the embedding gather and both eh RMSNorms run as one kernel on the replicated embedding table, bitwise equal to SGLang's;
  • eh_proj runs as a column-sharded GEMV, gathered by the target's all-reduce hop;
  • the top-1 is computed distributed instead of through the logits all-gather (up to 8 rows), and returns the one-hot LogitsProcessorOutput that SGLang's topk-1 draft path reads;
  • the draft layer's MLA prologue runs the target's attention kernels (up to 16 rows).

Other row counts run SGLang's code.

The draft spec requires two things:

5 draft steps at TP 8 (6 draft tokens)

  • Served draft lengths: --speculative-num-draft-tokens = --speculative-num-steps + 1, at 4 or 6. Any other length stops at argument resolution.
  • Widths: at T 6, verify steps of 1–2 requests (6 / 12 rows) run the kernels, and steps of 3–4 requests run SGLang's forward.
  • One-request steps: the split MoE fold hop, the dense-layer hops and the 16-warp sparse MLA decode now serve up to 6 rows, so one-request steps at T 6 use the same fast forms as at T 4.
  • No kernel source changed, and every T 4 width, MoE row map and kernel form is unchanged.

SGLang compatibility

Other files

  • plugin.py: two rows in SPECS.
  • README.md: the model and draft rows, slot table, requirements and launch command.
  • NOTICE.
  • .github/CODEOWNERS: one line.

Tests in this PR

  • test/unit/test_contract.py: adds the DSA binding view members, the TOLERATED_DSA_* lists and the GLM-5.3 mirror pins. These are skipped, with a named reason, on an SGLang without the binding view.
  • test/unit/test_glm5_3_glm5_2_config.py (CPU): runs the published nvidia/GLM-5.2-NVFP4 config through SGLang's requirement check for both specs. It also checks that any changed or missing pinned field is refused.
  • test/unit/test_glm5_3_rubin.py (CPU): checks that both specs admit SM107 and multi-node TP, and checks the no-multicast opt-in.

The GLM-5.3 kernel tests and the 5-draft-step unit tests follow in a separate PR; their results are below.

Accuracy Tests

8 x B300, TP 8, MTP (EAGLE, topk 1).

Parity. In a test-only parity mode, every eager verify step of a covered width runs SGLang's forward and then the plugin's stock step, compared on all 8 ranks.

Draft steps Model Verify steps Draft steps Max abs difference
3 GLM-5.3 560 at 4 / 8 / 12 / 16 rows 1,680 0
5 GLM-5.3 336 at 6 / 12 rows 2,760 0
5 GLM-5.2 280 at 6 / 12 rows 2,320 0

GSM8K-200 (5-shot, greedy), current branch. The SGLang base is main 9e9296fdf3 plus sgl-project/sglang#46210, the EAGLE step-graph PR (sgl-project/sglang#46211) and sgl-project/sglang#45795. The flags are --enable-nvidia-megakernel --enable-speculative-step-cuda-graph --hicache-async-ready-sync. The target and draft specs were active on 8 / 8 ranks.

Model Draft steps Concurrency Megakernel on Stock
GLM-5.3 3 4 0.995 —
GLM-5.3 5 4 0.990 —
GLM-5.3 (main 286e018594) 3 4 / 8 0.985 / 0.985 0.990 / 0.995
GLM-5.3 (main 286e018594) 5 4 0.985 0.995
GLM-5.2 (main 286e018594) 5 4 0.945 0.960

Full GSM8K test set (1,319 questions, 4 concurrent requests), GLM-5.3:

  • 5 draft steps: 97.27 % with the flag, 97.50 % without (McNemar p = 0.58).
  • 3 draft steps: 97.50 % with the flag, 97.57 % without (p = 1.0).

Other checks (3 draft steps, GLM-5.3, on earlier revisions of this PR with the same kernels):

  • KL of covered steps against the prefill path: mean 0.0287 (bound 0.03).
  • Passkey retrieval past the indexer's top-k span: found.
  • GSM8K-200 by slot set: attention,moe,hops 0.990; adding qkv_a_small_m_shared 0.995; adding dnorm 0.990; all six slots 0.995.
  • GLM-5.2 at 3 draft steps: GSM8K-200 0.975 with the megakernel, 0.960 without.

Acceptance length at 5 draft steps on GLM-5.3. In two pairs of GSM8K-200 runs, mean acceptance length was lower with the flag:

  • first pair: 3.710 against 4.162;
  • second pair: 3.989 against 4.418.

Most of the gap comes from one question (test index 119), which ran to the 16,384-token output limit with the flag. With 3–4 requests in flight, the servers were level in the second pair (4.37 vs 4.41).

With one request at a time, 5 draft steps reproduced the tokens of 3 draft steps up to 2,048 tokens of context. Per-request acceptance was 4.336 with the flag and 4.340 without.

GLM-5.2 (5.273 vs 5.182) did not show the gap. Acceptance-length parity with stock is not claimed for GLM-5.3 at 5 draft steps.

Kernel tests (GLM-5.3 test branch on this branch): 10 / 10 files pass.

  • MoE engines: match FlashInfer's TRT-LLM MoE bit for bit at every served row map, including 6 rows as (4, 2) and 12 rows as (4, 4, 4). Every map of the same rows gives the same bytes.
  • Attention, hop and draft kernels: within spec.NUMERICS and draft.NUMERICS.
  • compute-sanitizer: memcheck, synccheck and initcheck pass on every kernel form, and racecheck passes on the attention, draft and hop kernels.
  • Repeated launches: 400 per kernel and row count gave bitwise-identical outputs.

Speed Tests and Profiling

Static decode (benchmark/static_table.py, GLM-5.3, fixed acceptance length 2.99, --max-running-requests 5, --cuda-graph-max-bs-decode 4). Each cell is the mean of two launches in palindrome order, two passes each, in tok/s per user. These runs used an earlier revision of this PR, measured before qkv_a_small_m_shared and dnorm became default slots.

Concurrency Off (plugin not installed) attention,moe,hops + draft + draft / off
1 385.0 569.5 580.3 1.507
2 331.3 446.6 452.2 1.365
3 292.8 355.1 354.8 1.212
4 270.6 303.6 305.7 1.130

At concurrency 5, past the largest width, the server runs its usual path. There, + draft / off was 0.997.

Agentic coding trace (semianalysisai/cc-traces-weka-062126, replayed as recorded, about one hour per run). Fixed acceptance length 2.99, HiCache write-back with a 270 GB host pool, one run per cell.

GLM-5.3, 3 draft steps, SGLang main 74c08f3dc5; "stock" is that main without the flag:

Client concurrency Metric Stock This PR + sgl-project/sglang#46210
1 p90 interactivity (tok/s per user) 352.11 505.05 (1.43x)
1 p90 end-to-end normalized interactivity (tok/s per user) 157.64 200.61 (1.27x)
1 p90 time to first token (s) 1.271 1.130 (0.89x)

GLM-5.2, SGLang main dba22e9b68; "all on" adds sgl-project/sglang#46210, sgl-project/sglang#46211 and sgl-project/sglang#45795 with their flags:

Client concurrency --max-running-requests Draft steps (fixed acceptance) p90 interactivity, stock / all on p90 e2e normalized interactivity, stock / all on p90 TTFT (s), stock / all on
1 1 5 (3.61) 370.37 / 529.10 (1.43x) 175.41 / 204.44 (1.17x) 1.158 / 1.035 (0.89x)
1 1 3 (2.99) 348.43 / 512.82 (1.47x) 162.46 / 198.22 (1.22x) 1.157 / 1.025 (0.89x)
2 4 5 (3.61) 313.48 / 378.79 (1.21x) 149.80 / 204.80 (1.37x) 0.749 / 0.736 (0.98x)
2 4 3 (2.99) 300.30 / 401.61 (1.34x) 171.47 / 204.82 (1.19x) 0.654 / 0.753 (1.15x)

Support for 5 draft steps is for serving the 5-step recipes with the megakernel. At these fixed acceptance lengths it gives no speed gain over 3 draft steps.

The derived weight copies leave 3,241,024 KV-cache tokens, against 3,722,304 for stock (−12.9 %).

Checklist

  • pre-commit run --all-files passes.
  • Tests under test/ cover the change. This PR updates test/unit/test_contract.py and adds test/unit/test_glm5_3_glm5_2_config.py and test/unit/test_glm5_3_rubin.py. The GLM-5.3 kernel tests and the 5-draft-step unit tests follow in a separate PR.
  • The SGLang commit or PR this was tested against is named above.
  • Accuracy and speed results are attached when kernel or forward code changes.
  • A new model adds one line to .github/CODEOWNERS that names its owner.

…arget 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.
@nvrohanv
nvrohanv force-pushed the glm5_3-megakernel-specs branch from ef1ead3 to ad0bf91 Compare October 25, 2026 18:44
@YAMY1234

Copy link
Copy Markdown
Collaborator

Rechecked the static table on 8×B300 TP8, GLM-5.3 NVFP4, with ISL8192/OSL1024 and simulated acceptance 2.99. Measured snapshots: mega ad0bf9103d, hooks 465a9dd5b7 + step 3ba3fd1d2e on SGLang base dba22e9b68.

Concurrency Off Target Target + draft Target + draft / off
1 391.9 587.0 595.8 1.520×
2 348.1 472.5 477.6 1.372×
3 318.1 382.6 387.9 1.219×
4 298.1 332.1 337.5 1.132×

Throughput is tok/s/user. Same static_table.py protocol and slot sets as the original table: attention,moe,hops, then +draft; embedding replication on, step CUDA graphs and HiCache off. Two fresh launches per configuration in palindrome order, two passes per launch; each cell averages four pass medians.

Relative gains closely match the original table. Absolute throughput is higher in both the baseline and enabled arms, by roughly 8–10% at c3/c4. Target-only and target+draft are effectively tied at c3/c4 within the observed pass variation.

Conflicts in NOTICE, README.md and test/unit/test_contract.py; each keeps
both sides:

- NOTICE: the GLM-5.3 paragraph, then the Qwen3.5 one.
- README.md: the Qwen3.5 rows and paragraphs, then the GLM-5.3 / GLM-5.2
  ones.
- test/unit/test_contract.py: one SGLANG_MIRRORS check, main's (its
  mirror_digest and its tuple pins); the GLM-5.3 SGLANG_READS checks stay.
…equest)

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.
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.
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).
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.
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.
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.
DENSE_MAX_ROWS is 6 since the 5-draft-step change; the module docstring
still said 4.
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.
wenscarl added a commit to wenscarl/sgl-nvidia-mega that referenced this pull request Oct 29, 2026
Port sgl-project#24's 5 / 1 / 6 support (3889557, 492aa0a,
77015e7) to the GLM-5 DSA specs, which served speculative_num_draft_tokens
4 / speculative_num_steps 3 only.

- eligibility: SERVED_TOKENS_PER_REQUEST = (4, 6) and MAX_ROWS = 16;
  supported_rows(T) gives every whole number of requests up to 16 rows
  (T 4: 4 / 8 / 12 / 16, T 6: 6 / 12). The executor builds only the widths
  of the server's speculative_num_draft_tokens; the spec and Tier 1 refuse
  any other value or a pair with draft tokens != steps + 1, naming 4 / 3
  and 6 / 5.
- MoE: moe_split groups rows by 4 with the remainder last (6 -> (4, 2) on
  the two-group engine, 12 -> (4, 4, 4) on the four-group engine).
- Draft: the eh_proj shard rows are the draft decode's 1-4 plus the
  covered widths (T 6: 1, 2, 3, 4, 6, 12).
- One request at 6 rows keeps the fast forms: split MoE fold hop
  (SPLIT_MAX_ROWS 6), plain hops on dense layers (DENSE_MAX_ROWS 6) and the
  16-warp sparse MLA up to 6 rows. Kernels unchanged.

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

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ithout NVLink multicast

Port the Rubin pieces of sgl-project#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>
wenscarl and others added 11 commits November 4, 2026 16:54
[Feat] [Rubin] glm5_3: serve SM107, TP8 across two trays, and nodes without NVLink multicast
…ndex 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>
…nel 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>
…(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>
…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>
… 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>
…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>
…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>
…oj, 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>
…tX 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>
…d 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>
mattteochen and others added 5 commits November 6, 2026 22:58
…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>
[Perf] [GB300] glm5_3: TP4 kernel set, PD decode role and the draft-extend row select
…d 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>
Resolve the conflicts with main (Muse Spark, Kimi K3 and the optional-reads
contract):

- CODEOWNERS, plugin.py: keep both the GLM-5.3 and the Muse Spark entries.
- README: keep both the GLM-5.3 and the Kimi K3 rows, notes and sections.
- test_contract.py: combine main's SGLANG_OPTIONAL_READS with the GLM-5.3
  DSA-view / draft-check skips (the per-pattern list is renamed to skipped so
  it no longer shadows optional), and subtract both main's preparation set
  and the binding file's public functions from the untolerated names.

Follow main's API changes in glm5_3:

- Define speculative_runtime_hooks = None on the four built objects
  (GlmMegakernelVerify, FusedDraft, NoDraftStep, ComposedBuiltDraft), as the
  contract test added in sgl-project#43 requires.
- Read ForwardMode.is_target_verify from forward_mode.py (where SGLang now
  defines it), falling back to forward_batch_info.py.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@nvrohanv
nvrohanv marked this pull request as ready for review November 10, 2026 22:04
@nvrohanv
nvrohanv requested a review from Oasis-Git as a code owner November 10, 2026 22:04
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>
wenscarl and others added 2 commits November 10, 2026 16:40
[Fix] glm5_3: pin the SGLang mirrors for main's SGLang (cf6133a)
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>
@nvrohanv nvrohanv changed the title [Feat] [B300] glm5_3: NVIDIA megakernel specs for GLM-5.3 NVFP4 MTP target verify and NextN draft on B300 TP8 [Feat] [NVIDIA] glm5_3: NVIDIA megakernel specs for GLM-5.3 / GLM-5.2 NVFP4 MTP target verify and NextN draft Nov 10, 2026
@YAMY1234

Copy link
Copy Markdown
Collaborator

All unittests passed

@YAMY1234
YAMY1234 merged commit e6c85c7 into sgl-project:main Nov 10, 2026
4 of 5 checks passed
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.

4 participants