FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9

FP8 matrix multiply was requested on a GPU without hardware FP8 support. torch._scaled_mm requires compute capability 8.9 or newer (Ada, Hopper, Blackwell); Ampere (sm_80/sm_86) has no FP8 tensor cores and cannot emulate them at useful speed. This entry explains how to confirm the cause, apply the fix, and separate it from adjacent fp8-serving failures.

Symptom

FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9

Root cause

FP8 matrix multiply was requested on a GPU without hardware FP8 support. torch._scaled_mm requires compute capability 8.

Recommended fix

fall back to BF16 on this hardware. Confirm what you have: torch.cuda.get_device_capability(), it must be >= (8, 9).

How Denpex helps

Denpex investigates FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9 using the evidence you provide or your connected workload collects. Earlier rank, host or application evidence is needed to distinguish an initiating failure from a downstream report.

The literal signature is "FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9". It is a infrastructure failure associated with FP8 serving with SGLang, FlashInfer, and vLLM. The line identifies the failing operation or subsystem, while the surrounding evidence decides whether it is the initiating fault or a downstream symptom.

FP8 matrix multiply was requested on a GPU without hardware FP8 support. torch._scaled_mm requires compute capability 8.9 or newer (Ada, Hopper, Blackwell); Ampere (sm_80/sm_86) has no FP8 tensor cores and cannot emulate them at useful speed. The failure becomes visible at this call site because the operation first requires the missing resource, valid state, healthy peer, or correct result. Earlier log lines and a known-good control carry more causal value than the final wrapper exception.

• The workload stops or loses forward progress after emitting "FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9".
• A retry on the same configuration reproduces the failure because the causal state has not changed.
• The outer framework exception can hide the rank, node, allocation, or dependency that failed first.
• Increasing timeouts or reducing workload size can suppress the symptom without correcting the cause.

• FP8 serving with SGLang, FlashInfer, and vLLM
• production-shaped multi-accelerator workloads
• containerized and bare-metal deployments of the same stack

Use this checklist to test the hypothesis against a small reproduction. No single line proves the root cause, so preserve the preceding events and compare one variable at a time.

• ✓Find the first occurrence of "FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9" and preserve at least 100 lines before it.
• ✓Identify which rank, node, device, or process emitted the first related warning.
• ✓this usually appears when a config or container built for H100 is run on A100. Check whether the model checkpoint itself is FP8-quantized, if so it needs requantizing to BF16/INT8, not just a flag change.
• ✓Repeat the same input after the targeted change and require the signature to disappear.
• ✓Resume from job start with BF16, or reschedule onto sm_89+ only after the control passes.

• FP8 matrix multiply was requested on a GPU without hardware FP8 support. torch._scaled_mm requires compute capability 8.9 or newer (Ada, Hopper, Blackwell); Ampere (sm_80/sm_86) has no FP8 tensor cores and cannot emulate them at useful speed.
• The decisive evidence is the first log line that precedes "FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9" and differs from a healthy run.
• A nearby failure remains a competing hypothesis until a control separates configuration, capacity, transport, and hardware causes.

FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9

Use this text as a lookup key in logs and upstream issue trackers. It is not presented as a captured customer log. Confirm the cause from your own preceding events, versions, configuration and the cited references.

fall back to BF16 on this hardware. Confirm what you have: torch.cuda.get_device_capability(), it must be >= (8, 9). This changes the condition in the causal diagnosis instead of hiding the outer exception. The repeated control proves ownership before recovery from job start with BF16, or reschedule onto sm_89+.

# Preserve evidence before restarting
rg -n -i 'error|exception|timeout|failed' <log-file>
nvidia-smi
python -m torch.utils.collect_env

# Find the exact signature in the complete log
rg -n -F -- "FP8 GEMM torch._scaled_mm only supported on CUDA capability >= 8.9" <log-file>

Adapt the snippet to your framework. The same pattern holds for PyTorch Lightning, Hugging Face Trainer, DeepSpeed, Megatron-LM, and vLLM training wrappers. Where the wrapper exposes a config flag (for examplelr_scheduler_type in Trainer), prefer the flag over the imperative API to keep the schedule declarative and reproducible.

Select the traceback or open the failed terminal, then run Denpex locally to see the initiating rank, collateral failures, exact fix, and verification command without uploading the log.

CUDA reports errors asynchronously, so the traceback usually points at whatever line synchronised next rather than the one at fault. The hub covers every common CUDA error and how to make it report honestly.