perf(dsa): cuDNN + FlashMLA kernels for GLM-5 sparse attention (model.dsa_backend) by samsja · Pull Request #3829 · PrimeIntellect-ai/prime-rl

samsja · 2026-10-04T00:28:37Z

adding cudnn backward kernel for dsa, glm5 full training

New model.dsa_backend flag: tilelang (default, old path) or cudnn_flashmla (FlashMLA forward + cuDNN backward).
GLM-5.3 SFT example: base.toml now uses cudnn_flashmla, and a new 16k-12-node.toml overlay runs 12 nodes at 16k (cp=1, no offloading, fullgraph, full AC).

Part	Old (`tilelang`)	New (`cudnn_flashmla`)	Kernel speedup (H200)	Status
Sparse attention forward	TileLang	FlashMLA `flash_mla_sparse_fwd`	1.9–3.6×	Hopper tested; Blackwell built in the wheel, untested
Sparse attention backward	TileLang	cuDNN `dsa_bwd_sm90`	3–4.7×	Hopper tested; Blackwell wired in, untested

Full GLM-5.3 (78 layers), H200, 16k.

Fake data (random init, balanced routing)

setup	nodes	tok/s per GPU	MFU
original config (cp=2, offloading), `tilelang`	8	465	22.3%
`16k-12-node.toml`, `tilelang`	12	701	33.6%
`16k-12-node.toml`, `cudnn_flashmla`	12	835	40.1%

Real weights + data (INTELLECT-3-SFT-10K math, 12 nodes)

`dsa_backend`	step	tok/s	MFU
`tilelang`	31.2s	50.4k	25.2%
`cudnn_flashmla`	24.2s	64.9k	32.5%

🤖 Generated with Claude Code

Note

Medium Risk
Changes sparse-attention forward/backward kernels and gradient numerics (LSE/index flattening); cudnn_flashmla needs SM90+ and extra deps, with SM100 paths noted as untested.

Overview
Adds model.dsa_backend (tilelang default, cudnn_flashmla) so GLM-5 DSA sparse MLA can use FlashMLA sparse prefill forward and cuDNN CuTe-DSL backward instead of TileLang for both directions. The trainer passes the setting through model_config into sparse_mla, which now takes a backend argument; autograd routes backward to the matching implementation and normalizes LSE to natural log across paths (TileLang converts at the boundary).

Packaging adds nvidia-cudnn-frontend on the GPU extra, a flash-mla optional extra (pinned x86_64 wheel), and lockfile updates.

GLM-5.3 SFT examples: base.toml sets dsa_backend = "cudnn_flashmla"; new 16k-12-node.toml overlay documents 12-node / 16k training without CP or offloading; README expanded for both setups.

^{Reviewed by Cursor Bugbot for commit 0894aa8. Bugbot is set up for automated code reviews on this repo. Configure here.}

Removes duplicate and unused entries from `pyproject.toml` (rev-infra P9). The set of installed packages does not change for any extra combination (checked below). | Removed | Reason | |---|---| | second `orjson>=3.10` in `dependencies` | Duplicate of `orjson>=3.11.0` a few lines above. | | `pandas>=2.0` | Nothing imports it: no `import pandas` / `pd.` in `src/`, `tools/`, `scripts/`, `tests/`, `examples/`, `skills/`, `packages/`. It was added for the nano bridge (7089ba1). It stays installed because `datasets` depends on it. | | `pyarrow>=21.0.0` | No direct use anywhere in those directories. The one parquet writer, `tools/convert_traces_to_hf_dataset.py`, calls `datasets.Dataset.to_parquet`, and `datasets` depends on `pyarrow` itself. It stays installed. | | `nixl-cu13==0.10.1` in the `disagg` extra | `disagg` already lists `nixl`, and the `[[tool.uv.dependency-metadata]]` override for `nixl` 0.10.1 requires exactly `nixl-cu13==0.10.1`. The `gpu` extra keeps its explicit pin, because `transports/weights/nixl/agent.py` imports `nixl_cu13` directly. | | `mamba-ssm` dependency group | Duplicate of the `gpu` extra entry, which was added with the native Nemotron-H model (#3426). Every user of the group also syncs `--extra gpu` or `--all-extras`, so I dropped `--group mamba-ssm` from `Dockerfile.cuda` (2 places), `scripts/docker-entrypoint.sh` and the `pre_run_command` in `configs/advanced/nemotron-3-super/swe.toml`. The `[tool.uv.sources]` git pin and the build dependencies for mamba-ssm stay. | Behavior change: - None for installs: same locked packages and versions. - `uv sync --group mamba-ssm` no longer works; use `--extra gpu`, which every repo caller already passes. - `pandas` and `pyarrow` lose their direct lower bounds. They are still pulled in by `datasets`. ## Checks - `uv lock`: resolves 564 packages, the same set as before (568 `[[package]]` entries before and after). `uv lock --check` passes. The `uv.lock` diff only touches prime-rl's own metadata. I reverted the reordering of `[options.exclude-newer-package]` that my uv version (0.12.10) introduced. - I ran `uv export --frozen --all-packages` and compared the package lists before and after, ignoring the `# via` comment lines. They are identical for: no extras, `--extra disagg`, `--extra gpu`, `--extra dashboard`, `--all-extras`, and `--all-extras --all-groups`. `--extra disagg` still yields `nixl==0.10.1` and `nixl-cu13==0.10.1`, and `--extra gpu` still yields `mamba-ssm` at the pinned rev. - No Python changes, so no unit tests are affected. GPU: not needed. The Docker image build in CI covers the Dockerfile and entrypoint edits. ## Conflicts Text conflicts only in `pyproject.toml` / `uv.lock`, as with any dependency PR. None of the open PRs that touch these files (e.g. #3829, #3803, #3767, #3660) edit the lines removed here. A re-lock after rebase is enough.  --- > [!NOTE] > **Low Risk** > Metadata and sync-flag cleanup only; lockfile exports are unchanged and runtime installs should match prior behavior. > > **Overview** > Cleans up **duplicate and redundant dependency declarations** in `pyproject.toml` and aligns `uv.lock`, Docker, and Slurm sync commands with the slimmer manifest. **No intended change to installed packages** for existing extra combinations (`gpu`, `disagg`, `--all-extras`, etc.). > > **Removed from core `dependencies`:** duplicate `orjson>=3.10`, unused direct pins on `pandas` and `pyarrow` (they still come in via `datasets`). **Removed from `disagg` extra:** explicit `nixl-cu13==0.10.1` because `nixl` metadata already pulls that version; the **`gpu` extra keeps** its explicit `nixl-cu13` pin for direct imports. **Dropped the `mamba-ssm` dependency group** as a duplicate of the `gpu` extra entry—`Dockerfile.cuda`, `scripts/docker-entrypoint.sh`, and the Nemotron SWE `pre_run_command` no longer pass `--group mamba-ssm`. > > **Workflow note:** `uv sync --group mamba-ssm` alone is no longer supported; use `--extra gpu` (or `--all-extras`), which callers already use. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit f0f2ff6. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup>

cursor

Stale Bugbot comment from a previous run.

samsja · 2026-10-05T04:08:28Z

+    if backend == "cudnn_flashmla":
+        arch = cudnn_backward_arch(q, kv)
+        assert arch is not None, "dsa_backend='cudnn_flashmla' requires an SM90/SM100/SM103 GPU and a single KV head"
+        return cudnn_sparse_mla_backward(q, kv, out, grad_out, indices, lse, sm_scale, arch)
+    return tilelang_sparse_mla_backward(q, kv, out, grad_out, indices, lse, sm_scale)


can wwe do if then else instead of if return here please ?

Done in cb55aa6.

samsja · 2026-10-05T04:08:45Z

+        arch = cudnn_backward_arch(q, kv)
+        assert arch is not None, "dsa_backend='cudnn_flashmla' requires an SM90/SM100/SM103 GPU and a single KV head"


lets remove this assert please

Done in cb55aa6.

samsja · 2026-10-05T04:09:19Z

    block_I: int = 64,
    num_stages: int = 2,
    threads: int = 256,
+    backend: str = "tilelang",


can we use literatal here ? also why is the backend predefined should it be rather just a non optiona config

Done in cb55aa6.

samsja · 2026-10-05T04:09:47Z

+    if backend == "cudnn_flashmla":
+        assert cudnn_backward_arch(q, kv) is not None, (
+            "dsa_backend='cudnn_flashmla' requires an SM90/SM100/SM103 GPU and a single KV head"
+        )
+        assert importlib.util.find_spec("flash_mla") is not None, (
+            "dsa_backend='cudnn_flashmla' requires FlashMLA; install it with `uv sync --extra flash-mla`"
+        )
+        if sm_scale is None:
+            sm_scale = dim_plus_tail_dim**-0.5
+        return flashmla_sparse_mla_forward(q, kv, indices, sm_scale, d_v)
+


same as before can we do if then else, can we also remove the assert

Done in cb55aa6.

cursor

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit cb55aa6. Configure here.}

cursor · 2026-10-05T04:47:29Z

+    assert lse.shape == (B, S, H)

+    if backend == "cudnn_flashmla":
+        dq, dkv = cudnn_sparse_mla_backward(q, kv, out, grad_out, indices, lse, sm_scale, cudnn_backward_arch(q, kv))


None arch selects SM100 backward

Medium Severity

cudnn_flashmla backward now passes cudnn_backward_arch() straight into cudnn_sparse_mla_backward. That helper returns None for a non-single KV head or an unsupported GPU, and the callee treats every non-sm90 value as SM100, so the wrong kernel runs instead of failing closed.

Additional Locations (1)

src/prime_rl/trainer/models/kernels/sparse_mla_bwd.py#L361-L371

^{Reviewed by Cursor Bugbot for commit cb55aa6. Configure here.}

…on SM90 The vendored TileLang sparse MLA backward reaches ~93-130 TFLOP/s on H200 (no pipelining, fp32 atomics for every top-k slot including sentinels). On Hopper, dispatch to cuDNN frontend's SM90 DSA backward instead (warp-specialized, TMA Q/dO, cp.async KV gather, skips dKV atomics for masked slots): 3.5-4.7x faster, ~410-440 TFLOP/s at S=8k/16k, topk=2048. Sentinel indices are remapped to -1 and the base-2 LSE is converted to natural log. TileLang remains the fallback on other architectures. Bumps nvidia-cudnn-frontend to >=1.30 (1.25 is incompatible with the pinned CUTLASS DSL 4.7.1) and adds a standalone microbenchmark with an fp32 reference check. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

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

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

…ward route Replace model.sparse_mla_backward with model.dsa_backend, selecting the DSA sparse MLA kernels for both passes through the single prime_rl::sparse_mla op: - tilelang (default): TileLang forward, cuDNN backward on SM90, TileLang backward elsewhere. - cudnn_flashmla: FlashMLA flash_mla_sparse_fwd forward + cuDNN backward (SM90; the SM100/SM103 cuDNN backward is wired up but untested). - tilelang_legacy: TileLang forward and backward. The op now always returns a natural-log LSE. The FlashMLA route flattens the batch, maps sentinel indices to -1 and passes a topk_length covering every valid index so the sentinel tail is skipped. FlashMLA (nv_dev @ b7643bd, sm_90a + sm_100f) ships as the new flash-mla extra (part of `all`) from a prebuilt torch 2.13 / cu130 x86_64 wheel. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

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

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

… dsa_backend=cudnn_flashmla On SM90 the cudnn_flashmla route now selects DSA indices with the cuDNN frontend SM90 FP8 indexer forward (ratio=1 token-level causal mask, packed documents as THD segments with q_causal_offsets for the CP shard) and the cuDNN radix top-k (per-row causal lengths, chunked to bound its scratch buffer), behind the prime_rl::cudnn_fp8_indexer custom op (AC mandatory save). Same UE8M0 FP8 quantization as the Triton indexer; ~2x faster per layer on H200. SM100 and the tilelang backend keep the Triton indexer. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…k tie behaviour Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…ent top-k tie behaviour" This reverts commit 06f5c60.

…-k under dsa_backend=cudnn_flashmla" This reverts commit 5724fef.

…atch without asserts Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…3968) Sets `dsa_backend = "cudnn_flashmla"` for the trainer in the GLM-5.3 RL example (`examples/advanced/glm-5.3/swe.toml`). The SFT example already uses it (#3829); the RL example still ran the TileLang forward and backward. Measured on 1x8 H200 with the trainer settings of `swe.toml` (131k seq, cp 4, ep 8, fp8 + `deepgemm_fp8`, full AC + offload, compile). Random init, 8 layers (3 dense + 5 MoE), fake data, 2 micro-batches per dp rank, balanced routing: | `dsa_backend` | step 2 / 3 | sparse attn fwd per call | sparse attn bwd per call | |---|---|---|---| | `tilelang` | 16.6 s / 15.5 s | 35 ms | 190 ms | | `cudnn_flashmla` | 12.8 s / 12.1 s | 12 ms | 53 ms | - Entropy and mismatch KL are identical to 4 decimals. - With a 4-layer run (5.7 s), one MoE layer costs 1.6 s instead of 2.07 s. Extrapolated to 78 layers at this shape: ~124 s vs ~160 s per step (-22%). - Kernel check at the CP shape (8k local queries vs 32k keys, cp rank 0 and 3, top-k 2048, sentinel slots): out / dq / dkv relative error vs TileLang 5e-4 to 9e-4, LSE 7e-8. - Needs the `flash-mla` extra, which the README's `uv sync --all-extras` installs.  --- > [!NOTE] > **Low Risk** > Single example TOML knob with no production code changes; requires the optional flash-mla dependency for the new backend. > > **Overview** > Aligns the GLM-5.3 RL example trainer config with the SFT example by setting **`dsa_backend = "cudnn_flashmla"`** in `examples/advanced/glm-5.3/swe.toml`, so sparse DSA attention uses the cuDNN/FlashMLA path instead of the default TileLang forward/backward. > > This is a **performance-only** example change for long-context RL training (131k seq, CP/EP, fp8 MoE); behavior should match TileLang within measured numerical tolerance when the `flash-mla` extra is installed. > > <sup>Reviewed by [Cursor Bugbot](https://cursor.com/bugbot) for commit e2135c3. Bugbot is set up for automated code reviews on this repo. Configure [here](https://www.cursor.com/dashboard/bugbot).</sup>

… example - examples/advanced/qwen3-235b-a22b/sft-max-perf.toml: 4 x 8 B300, seq 16k, EP 8, full AC, fullgraph, fp32 AdamW, no offloading, balanced routing. Measured 4.39 s/step, 46.6% MFU. Ported from #3765 to the current schema (no `model.impl`, `offload = "none"`, no cluster-specific paths). - docs/performance-trainer.md: how MFU is computed, the full setup and numbers for the Qwen3-235B run, and the GLM-5.3 SFT before/after for the cuDNN + FlashMLA sparse attention backend (#3829). - Linked from docs nav, docs/scaling.md#benchmarking, and examples/README.md. Co-Authored-By: Mario Sieg <mario.sieg.64@gmail.com> Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

…forward, cuDNN backward) (#4078) Adds a `cudnn_flashmla` option to `model.dsa_backend` for DeepSeek V4 that runs the same FlashMLA sparse prefill forward vLLM uses for DeepSeek V4 prefill, so the trainer and inference compute attention with the same kernel. The backward is the cuDNN frontend's DSA kernel; it runs on SM90, SM100 and SM103, and `tilelang` stays the default. - `cudnn_dsa_patch` stops cuDNN recompiling its backward each time sequence length or packing changes. - On B300 (SM103), both DSv4 test files pass and the kernels match a float32 reference; no end-to-end run there yet. - On H200 at the B300 changes (`6a1cae560`), both DSv4 test files pass again, and a 3-step end-to-end run (cp 8, ep 8) compiles 3 cuDNN kernels per rank in step 1 and matches the earlier losses, step time and peak memory. - Fixes GLM-5's SM100 `cudnn_flashmla` backward, which omitted cuDNN's required workspace and always raised. - `tilelang` stays the shared `dsa_backend` default and `configs/advanced/deepseek-v4-flash/sft.toml` opts in to `cudnn_flashmla`, following #3829, which did the same for the GLM-5.3 configs. FlashMLA only ships an x86_64 wheel in an optional extra, so a `cudnn_flashmla` default would break aarch64 and installs without the extra. We can flip the default instead if reviewers prefer. DeepSeek-V4-Flash first 6 layers, real weights, cp 8, ep 8, one 8xH200 node. Sparse attention GPU time per step, rank 0 profiler trace at 8k tokens/GPU: | kernels | tilelang | cudnn_flashmla | speedup | |---|---|---|---| | forward, incl. recompute (48 calls) | 56.3 ms | 26.5 ms | 2.12x | | backward (24 calls) | 71.3 ms | 36.7 ms | 1.94x | | all sparse attention | 136.0 ms | 76.7 ms | 1.77x | Sparse attention is under 5% of the step here, so the end-to-end gain is smaller. Median steady-state step time: | tokens/GPU | tilelang | cudnn_flashmla | speedup | max loss diff vs tilelang | max loss diff, two tilelang runs | |---|---|---|---|---|---| | 8k | 2.86 s | 2.80 s | 1.02x | 1.0e-3 | 1.7e-3 | | 16k | 4.59 s | 4.36 s | 1.05x | 7.3e-4 | 1.4e-3 | Peak memory is equal, and each rank compiles 3 cuDNN kernels, all in step 1. Baseline [`main`](7c9d474); [results](https://github.com/PrimeIntellect-ai/prime-rl/blob/524ec668e5684134210cef64b3575476fdf59693/benchmarks/scripts/dsv4_cudnn_flashmla_pr/results/README.md) and [config](https://github.com/PrimeIntellect-ai/prime-rl/blob/524ec668e5684134210cef64b3575476fdf59693/benchmarks/scripts/dsv4_cudnn_flashmla_pr/sft.toml) at `524ec668e`. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com>

faresobeid mentioned this pull request Oct 4, 2026

chore(deps): drop duplicate and unused pyproject dependencies #3856

Merged

samsja mentioned this pull request Oct 5, 2026

feat(trainer): DeepSeek-V4.1 Flash (engram, shared compressed attention, CP) #3879

Closed

samsja changed the title ~~perf(kernels): use cuDNN CuTe-DSL DSA kernel for sparse MLA backward …~~ perf(dsa): cuDNN + FlashMLA kernels for GLM-5 sparse attention (model.dsa_backend) Oct 5, 2026

cursor Bot reviewed Oct 5, 2026

View reviewed changes

Comment thread src/prime_rl/trainer/models/kernels/cudnn_indexer.py Resolved

Comment thread src/prime_rl/trainer/models/kernels/cudnn_indexer.py Resolved

samsja mentioned this pull request Oct 5, 2026

perf(dsa): cuDNN indexer + radix top-k for dsa_backend=cudnn_flashmla #3883

Closed

samsja commented Oct 5, 2026

View reviewed changes

JohannesHa previously approved these changes Oct 5, 2026

View reviewed changes

samsja dismissed JohannesHa’s stale review via cb55aa6 October 5, 2026 04:44

cursor Bot reviewed Oct 5, 2026

View reviewed changes

samsja and others added 11 commits October 5, 2026 04:51

chore: drop sparse MLA microbenchmark script

30cd8c9

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

feat(config): model.sparse_mla_backward to choose the DSA backward ke…

528f731

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

examples(glm-5.3): SFT on cudnn_flashmla DSA backend, add 12-node 16k…

1b935b6

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

refactor(trainer): dsa_backend is tilelang (TileLang fwd+bwd) or cudn…

f2372e9

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

docs(glm-5.3): note the cuDNN indexer on cudnn_flashmla; comment top-…

513ccfc

…k tie behaviour Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Revert "docs(glm-5.3): note the cuDNN indexer on cudnn_flashmla; comm…

0388753

…ent top-k tie behaviour" This reverts commit 06f5c60.

Revert "perf(trainer): cuDNN frontend FP8 indexer scoring + radix top…

fb5da5b

…-k under dsa_backend=cudnn_flashmla" This reverts commit 5724fef.

refactor(dsa): DSABackend Literal, required backend arg, if/else disp…

0894aa8

…atch without asserts Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

samsja force-pushed the perf/sparse-mla-bwd branch from cb55aa6 to 0894aa8 Compare October 5, 2026 04:52

samsja merged commit 2b688ed into main Oct 5, 2026
18 of 20 checks passed

samsja deleted the perf/sparse-mla-bwd branch October 5, 2026 04:54

faresobeid mentioned this pull request Oct 10, 2026

perf(examples): use cudnn_flashmla DSA backend in GLM-5.3 RL example #3968

Merged

samsja mentioned this pull request Oct 12, 2026

docs(scaling): max-throughput recipe + Qwen3-235B-A22B SFT max trainer perf example #4046

Merged

garrett361 mentioned this pull request Oct 14, 2026

feat(deepseek-v4): cudnn_flashmla sparse attention backend (FlashMLA forward, cuDNN backward) #4078

Merged

Navigation Menu

Repository navigation

perf(dsa): cuDNN + FlashMLA kernels for GLM-5 sparse attention (model.dsa_backend) - #3829

perf(dsa): cuDNN + FlashMLA kernels for GLM-5 sparse attention (model.dsa_backend)#3829
samsja merged 11 commits into
mainfrom
perf/sparse-mla-bwd

samsja commented Oct 4, 2026 •

edited by cursor Bot

Loading

Uh oh!

cursor Bot left a comment •

edited

Loading

Uh oh!

Uh oh!

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

samsja Oct 5, 2026

Uh oh!

cursor Bot left a comment

Uh oh!

cursor Bot Oct 5, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		arch = cudnn_backward_arch(q, kv)
		assert arch is not None, "dsa_backend='cudnn_flashmla' requires an SM90/SM100/SM103 GPU and a single KV head"

Conversation

samsja commented Oct 4, 2026 • edited by cursor Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cursor Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cursor Bot left a comment

Choose a reason for hiding this comment

Uh oh!

cursor Bot Oct 5, 2026

Choose a reason for hiding this comment

None arch selects SM100 backward

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

samsja commented Oct 4, 2026 •

edited by cursor Bot

Loading

cursor Bot left a comment •

edited

Loading