Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦
C
deepseek_v4: say why no CUDA backend DLL was usable (#1906)
JustVugg78e98d16ceab
diff --git a/c/backend_loader_dsv4.c b/c/backend_loader_dsv4.cindex 4ce09ef1..ec512eee 100644--- a/c/backend_loader_dsv4.c+++ b/c/backend_loader_dsv4.c@@ -376,15 +376,30 @@ static int dsv4_cuda_load(void){int n = 0;if(forced && *forced) candidates[n++] = forced;else { candidates[n++] = DSV4_BACKEND_DLL_DG; candidates[n++] = DSV4_BACKEND_DLL; }+ /* why each candidate was passed over, for the closing line (#1906: it named no cause) */+ char why[512]; size_t wn = 0; why[0] = 0;for(int i = 0; i < n; i++){g_dsv4.dll = dsv4_load_one(candidates[i]);- if(!g_dsv4.dll) continue;- if(!dsv4_cuda_resolve(candidates[i])) continue; /* unloaded inside */+ if(!g_dsv4.dll){+ DWORD err = GetLastError();+ int w = snprintf(why + wn, sizeof(why) - wn, "%s%s: %s (Windows error %lu)", wn ? "; " : "", candidates[i],+ err == ERROR_MOD_NOT_FOUND ? "not next to the engine, or a DLL it needs (the CUDA runtime) is missing"+ : err == ERROR_BAD_EXE_FORMAT ? "not a 64-bit DLL" : "not loadable", (unsigned long)err);+ if(w > 0 && (size_t)w < sizeof(why) - wn) wn += (size_t)w;+ continue;+ }+ if(!dsv4_cuda_resolve(candidates[i])){ /* unloaded inside; it said which symbol */+ int w = snprintf(why + wn, sizeof(why) - wn, "%s%s: a missing symbol", wn ? "; " : "", candidates[i]);+ if(w > 0 && (size_t)w < sizeof(why) - wn) wn += (size_t)w;+ continue;+ }/* Optional compatibility probe: a build whose kernels cannot run on* device 0 says so, and the next candidate is tried. */if(g_dsv4.backend_arch_ok && !g_dsv4.backend_arch_ok(0)){fprintf(stderr, DSV4_VENDOR_TAG " %s does not fit this GPU; trying the next backend\n",candidates[i]);+ int w = snprintf(why + wn, sizeof(why) - wn, "%s%s: not built for this GPU", wn ? "; " : "", candidates[i]);+ if(w > 0 && (size_t)w < sizeof(why) - wn) wn += (size_t)w;FreeLibrary(g_dsv4.dll); g_dsv4.dll = NULL;memset(&g_dsv4, 0, sizeof(g_dsv4)); g_dsv4.loaded = 1;continue;@@ -395,8 +410,10 @@ static int dsv4_cuda_load(void){g_dsv4.available = 1;return 1;}- fprintf(stderr, DSV4_VENDOR_TAG " no usable backend DLL (" DSV4_BACKEND_DLL_DG " / "- DSV4_BACKEND_DLL "); GPU tier disabled (CPU path remains active).\n");+ fprintf(stderr, DSV4_VENDOR_TAG " no usable backend DLL (%s); GPU tier disabled (CPU path remains active). "+ "The DLLs need an sm_80+ GPU (Ampere or newer); Pascal and Turing take a build with "+ "CUDA_ARCH=portable-pre-ampere NO_TC=1 (docs/deepseek-v4.md).\n",+ wn ? why : DSV4_BACKEND_DLL_DG " / " DSV4_BACKEND_DLL);return 0;}