/* Write a tiny model fixture to a directory. * * Thin command-line front end over tests/model_fixture.h, so out-of-process * callers (shell scripts, Python) can materialize the same synthetic model the * tests build in memory. The weights are deterministic but meaningless: the * fixture exercises loading and CLI/server plumbing, not embedding quality. * * Usage: gen_fixture --out DIR [--dtype F32|BF16|F16] [--model base|qwen3|qwen2] * [--hidden N] [--heads N] [--kv-heads N] [--head-dim N] * [--intermediate N] [--vocab N] [--eos-id N] [--no-vocab] * [--late-dim N] * * --late-dim N also writes a 1_Dense/linear.weight projection (token_dim N over * the hidden size), turning the fixture into a late-interaction/rerank model. */ #include "model_fixture.h" #include #include #include static void usage(FILE *f) { fputs("usage: gen_fixture --out DIR [--dtype F32|BF16|F16]\n" " [--model base|qwen3|qwen2]\n" " [--hidden N] [--heads N] [--kv-heads N] [--head-dim N]\n" " [--intermediate N] [--vocab N] [--eos-id N] [--no-vocab]\n" " [--late-dim N]\n", f); } int main(int argc, char **argv) { const char *out = NULL; const char *dtype = "F32"; mf_model_kind_t model = MF_MODEL_BASE; int eos_id = TF_EOT_ID; int write_vocab = 1; int late_dim = 0; tm_dims_t d = mf_default_dims(); for (int i = 1; i < argc; i++) { const char *a = argv[i]; if (!strcmp(a, "-h") || !strcmp(a, "--help")) { usage(stdout); return 0; } else if (!strcmp(a, "--out") && i + 1 < argc) { out = argv[++i]; } else if (!strcmp(a, "--dtype") && i + 1 < argc) { dtype = argv[++i]; } else if (!strcmp(a, "--model") && i + 1 < argc) { if (mf_parse_model_kind(argv[++i], &model) != 0) { fprintf(stderr, "gen_fixture: --model must be base, qwen3, or qwen2\n"); return 2; } } else if (!strcmp(a, "--hidden") && i + 1 < argc) { d.hidden = atoi(argv[++i]); } else if (!strcmp(a, "--heads") && i + 1 < argc) { d.heads = atoi(argv[++i]); } else if (!strcmp(a, "--kv-heads") && i + 1 < argc) { d.kv_heads = atoi(argv[++i]); } else if (!strcmp(a, "--head-dim") && i + 1 < argc) { d.head_dim = atoi(argv[++i]); } else if (!strcmp(a, "--intermediate") && i + 1 < argc) { d.intermediate = atoi(argv[++i]); } else if (!strcmp(a, "--vocab") && i + 1 < argc) { d.vocab = atoi(argv[++i]); } else if (!strcmp(a, "--eos-id") && i + 1 < argc) { eos_id = atoi(argv[++i]); } else if (!strcmp(a, "--late-dim") && i + 1 < argc) { late_dim = atoi(argv[++i]); } else if (!strcmp(a, "--no-vocab")) { write_vocab = 0; } else { fprintf(stderr, "gen_fixture: unknown or incomplete option: %s\n", a); usage(stderr); return 2; } } if (!out) { fprintf(stderr, "gen_fixture: --out DIR is required\n"); usage(stderr); return 2; } if (!mf_valid_dtype(dtype)) { fprintf(stderr, "gen_fixture: --dtype must be F32, BF16, or F16\n"); return 2; } if (mf_write_fixture(out, dtype, &d, model, eos_id, write_vocab, /*qwen2_zero_bias=*/0) != 0) { fprintf(stderr, "gen_fixture: failed to write fixture to %s\n", out); return 1; } /* A late-interaction/rerank model adds a Dense projection over the hidden * size; the base weights and vocab above are otherwise unchanged. */ if (late_dim > 0 && tm_write_late_projection(out, dtype, late_dim, d.hidden) != 0) { fprintf(stderr, "gen_fixture: failed to write late projection to %s\n", out); return 1; } printf("%s\n", out); return 0; }