cuda: guard PDL can-use check on pre-Hopper devices (fix sm_86 crash)

[ddvnguyen]

Overview

ggml_cuda_kernel_can_use_pdl() (ggml/src/ggml-cuda/common.cuh) called cudaFuncGetAttributes() on every device before checking attr.ptxVersion >= 90. On sm_86 (RTX 3060) + CUDA 13.2 that query faults for the PDL-carrying flash-attn kernels, and CUDA_CHECK escalates the error into a GGML_ABORT — crashing rpc-server on the 3060 peer.

This is the GGML_CUDA_PDL crash tracked in PR #31's live-hardware notes ("rpc-server crashes on 3060 with PDL check"), and it is the blocker that forces COMBINE onto the --rpc-engine add-backend-after-load path (→ the 35-split perf gap) instead of the clean native --rpc register-before-load path.

Fix

Gate on device compute capability before the cudaFuncGetAttributes query and cache a permanent false for cc < GGML_CUDA_CC_HOPPER. Pre-Hopper GPUs never emit PDL device-side primitives (ggml_cuda_pdl_sync/_lc are compiled out for arch < Hopper), so the query is pointless there anyway. Hopper/Blackwell (including sm_120 / the 5060 Ti) are unaffected — they still run the full PTX-version check.

Additional information

• Root cause and hardware repro documented in ddvnguyen/llama.cpp#31 (comments of 2026-07-06).
• Unblocks the native --rpc register-before-load architecture that PR fork: resolve #376 — make COMBINE mode RPC work (gate-then-degrade) #31 needs to collapse the 35+ graph-split regression.
• ⚠️ Not compiled in the authoring environment (no CUDA / sm_86 host). Needs a hardware build to confirm the crash is gone: cmake --build build_sm86_sm120 --target llama-engine, then run rpc-server on the 3060 and exercise a flash-attn graph.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES — change drafted by an AI agent (Claude Code) from the PR fork: resolve #376 — make COMBINE mode RPC work (gate-then-degrade) #31 crash analysis; author is responsible for reviewing, building, and hardware-verifying before merge. This project restricts AI-generated content — see AGENTS.md and CONTRIBUTING.md.

---

Generated by Claude Code

[ddvnguyen]

Closing — this guard rests on a misdiagnosis.

Spike hydra_vortex#381 ran the --rpc + --tensor-split path with rpc-server on the 3060 (sm_86), -fa 1, on fork 8d18d1b12 — which already contained the full PDL subsystem (git show 8d18d1b:ggml/src/ggml-cuda/common.cuh has the identical PDL block). Flash-attn ran on ~5 attention layers on the 3060 and it worked (620 prefill / 20.3 decode, no crash). Since the main flash-attn kernel launches through ggml_cuda_kernel_launch → ggml_cuda_kernel_can_use_pdl → cudaFuncGetAttributes (fattn-common.cuh:1207), spike ggml-org#381 is empirical proof that cudaFuncGetAttributes on sm_86 does not fault.

So there is no PDL crash to fix here. And pre-Hopper already returns false via the existing attr.ptxVersion >= 90 check, making this guard behaviorally inert. PR #31's "rpc-server crashes on 3060 with PDL check" is a misdiagnosis — the real COMBINE blocker is -ot/expert-split recurrent-layer state corruption plus add-RPC-after-load graph splits, both addressed by the layer-split path (ggml-org#383). Superseded by that work.

---

Generated by Claude Code