[sergqwer]

Summary

• A clear stop for a wrong build. Before the expert arena loads, the engine names the card and its compute
  capability, and checks that the build has code for it (cudaFuncGetAttributes on a kernel). A binary built for
  other GPUs used to fail at its first kernel, after the whole arena had loaded. Now it stops in ~0.3 s, names the
  card and its sm_, and says to rebuild (setup does).
• STRATA_EMULATE_CC=75|80|86|89, for tests. With a PTX build for an older generation (e.g.
  CMAKE_CUDA_ARCHITECTURES=86-virtual, JIT-compiled on a newer card), the host's device queries answer as that
  card: compute capability for ggml's MMQ configs and the QSA kernel choice, shared memory per block (64 KB on
  Turing). The kernels are then the real card's, so their results are too, though their speed is not. This is how
  the fork tested RTX 20/30/40 code paths on an RTX 5090.
• STRATA_QSA_WARP=1|select|attn: the pre-sm_80 QSA kernels on any card, for A/B.

Measured

On the fork, an sm_86-only build on the RTX 5090 stops in 0.3 s with the card's name. Emulated sm_75/86/89 builds
give first-token KL to the native sm_120a build of 0.0005 (sm_89), 0.0003 (sm_86) and 0.0034 (sm_75), all with the
same top-1. On main (IQ2_XS) the check prints the card and changes nothing else.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VZy1yKaDDiA8a7svdwaHio

[Niko1221]

Thanks! This is in the 0.1.32 batch. We take it with a small fix (a placeholder name when the device query fails); every change goes through our identity gate (byte-identical output on four models) and a speed A/B before it ships.

[sergqwer]

Two things from rebuilding our fork on 0.1.31 with this PR applied.

1. 0.1.31's new sm_75 tile in fused_gr_read_multi reads the real compute capability. cudaDeviceGetAttribute there decides small_tile, so under STRATA_EMULATE_CC=75 the host picks the 2560-float tile while the device code is Turing's. That breaks the rule the emulation rests on: the host must answer as the emulated card. In our release:

   const bool small_tile = strata::cc_major_of(cc_maj) * 10 + strata::cc_minor_of(cc_min) == 75;

2. device_info() can answer as the emulated card too. Our fork does it:

   d.cc_major = cc_major_of(p.major);
   d.cc_minor = cc_minor_of(p.minor);

   It also prints one EMULATING compute capability line, so anything that sizes from DeviceInfo agrees with the kernels. It is not in this PR; add it if you take the emulation.

The card check stays before the expert arena starts loading. On 0.1.31 that is right before the expert sources are opened.

On 0.1.31: sergqwer/strata-nvfp4@6b5954a

🤖 Generated with Claude Code

[Niko1221]

Released in 0.1.32 with a placeholder name when the device query fails. Merged under your name (GitHub doesn't show it as merged because it went in by hand). Thank you!