Run frontier MoE models on hardware you already own β pure C, zero deps, experts streamed from disk. Tiny engine, immense model. π¦
C
Fix duplicate entries in deepseek-v4 documentation
Muhammad Jamshaid Iqbal75acdbe90596
diff --git a/docs/deepseek-v4.md b/docs/deepseek-v4.mdindex 0450b97d..30a2b0df 100644--- a/docs/deepseek-v4.md+++ b/docs/deepseek-v4.md@@ -302,15 +302,14 @@ is no software path to fall back to.| CPU only (no DLL / `DSV4_CUDA=0` / Linux default) | β | CPU reference | CPU reference || generic DLL / `CUDA=1` | any sm_80+ | GPU attention block, indexer, generic batched MoE on the VRAM bank | GPU attention, indexer, expert mirrors || generic, pre-Ampere / `CUDA=1 CUDA_ARCH=portable-pre-ampere NO_TC=1` | sm_61 (Pascal), sm_75 (Turing) | as generic | as generic |+| DeepGEMM DLL / `CUDA=1 DEEPGEMM=1` | compute 12.x | as generic + tensor-core dense/MoE GEMMs | same as generic |+| multi-GPU | β | single device today (`DSV4_CUDA_DEVICE` selects); expert-parallel design drafted, not implemented | β |The runtime check `dsv4_cuda_backend_arch_ok` admits sm_60 and up for thegeneric build. It is deliberately independent of what the binary contains: a`CUDA_ARCH=portable` build has no sm_61/sm_75 cubin, and running it on such a-card fails at launch with "no kernel image is available" rather than producing+card fails at launch with `"no kernel image is available"` rather than producinga wrong answer. Build with `portable-pre-ampere` for those cards.-| DeepGEMM DLL / `CUDA=1 DEEPGEMM=1` | compute 12.x | as generic + tensor-core dense/MoE GEMMs | same as generic |-| multi-GPU | β | single device today (`DSV4_CUDA_DEVICE` selects); expert-parallel design drafted, not implemented | β |-## Environment reference (V4 engine)Defaults in parentheses; all read by `c/deepseek_v4.c` unless noted `.cu`.