diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-17 01:21:41 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-17 01:21:41 +0200 |
| commit | e00becb643b3a99de16462b3a032717a93b983d9 (patch) | |
| tree | 873ad2646b0ef2f07eae5a5a74387d062c77ec0c /tests/test_bf16_model.c | |
| parent | b4343d6ac1c6cc7a7d228f954220def2f6e6ed67 (diff) | |
| download | ffwd-e00becb643b3a99de16462b3a032717a93b983d9.tar.gz ffwd-e00becb643b3a99de16462b3a032717a93b983d9.zip | |
all: accept F16 safetensors weights (widen to F32 at load)
Diffstat (limited to 'tests/test_bf16_model.c')
| -rw-r--r-- | tests/test_bf16_model.c | 59 |
1 files changed, 39 insertions, 20 deletions
diff --git a/tests/test_bf16_model.c b/tests/test_bf16_model.c index 48e4766..0849f3e 100644 --- a/tests/test_bf16_model.c +++ b/tests/test_bf16_model.c @@ -1,6 +1,7 @@ -/* tests/test_bf16_model.c - F32 vs BF16 loader/dispatch parity on a tiny - * synthetic model. Hermetic: builds its own model dirs. Runs via - * `make test`. */ +/* tests/test_bf16_model.c - F32 vs BF16 vs F16 loader/dispatch parity on a tiny + * synthetic model. Hermetic: builds its own model dirs. Runs via `make test`. + * F16 has no fused kernel: the loader widens it to F32 once at load, so an F16 + * copy must match the F32 copy as closely as the BF16 copy does. */ #include "embed.h" #include "tiny_model.h" @@ -12,7 +13,7 @@ #include <unistd.h> int main(void) { - char root[1024], f32_dir[1088], bf16_dir[1088], prefixed_dir[1088]; + char root[1024], f32_dir[1088], bf16_dir[1088], f16_dir[1088], prefixed_dir[1088]; snprintf(root, sizeof(root), "%s/embed-bf16-test-XXXXXX", getenv("TMPDIR") ? getenv("TMPDIR") : "/tmp"); if (!mkdtemp(root)) { @@ -21,10 +22,12 @@ int main(void) { } snprintf(f32_dir, sizeof(f32_dir), "%s/f32", root); snprintf(bf16_dir, sizeof(bf16_dir), "%s/bf16", root); + snprintf(f16_dir, sizeof(f16_dir), "%s/f16", root); snprintf(prefixed_dir, sizeof(prefixed_dir), "%s/prefixed", root); static const tm_dims_t dims = {4, 2, 1, 2, 8, 16}; - if (mkdir(f32_dir, 0755) != 0 || mkdir(bf16_dir, 0755) != 0 || mkdir(prefixed_dir, 0755) != 0 || - tm_write_model(f32_dir, "F32") != 0 || tm_write_model(bf16_dir, "BF16") != 0 || + if (mkdir(f32_dir, 0755) != 0 || mkdir(bf16_dir, 0755) != 0 || mkdir(f16_dir, 0755) != 0 || + mkdir(prefixed_dir, 0755) != 0 || tm_write_model(f32_dir, "F32") != 0 || + tm_write_model(bf16_dir, "BF16") != 0 || tm_write_model(f16_dir, "F16") != 0 || tm_write_prefixed_model_dims(prefixed_dir, "F32", &dims) != 0) { fprintf(stderr, "fixture creation failed\n"); return 2; @@ -33,8 +36,9 @@ int main(void) { int ids[] = {1, 2, 3, 4}; embed_model_t *mf32 = embed_model_load(f32_dir); embed_model_t *mbf16 = embed_model_load(bf16_dir); + embed_model_t *mf16 = embed_model_load(f16_dir); embed_model_t *mprefixed = embed_model_load(prefixed_dir); - if (!mf32 || !mbf16 || !mprefixed) { + if (!mf32 || !mbf16 || !mf16 || !mprefixed) { fprintf(stderr, "model load failed\n"); return 1; } @@ -44,32 +48,39 @@ int main(void) { float *a = (float *)malloc((size_t)dim * sizeof(float)); float *b = (float *)malloc((size_t)dim * sizeof(float)); float *c = (float *)malloc((size_t)dim * sizeof(float)); + float *e = (float *)malloc((size_t)dim * sizeof(float)); embed_workspace_t *wf32 = embed_workspace_new(mf32); embed_workspace_t *wbf16 = embed_workspace_new(mbf16); + embed_workspace_t *wf16 = embed_workspace_new(mf16); embed_workspace_t *wprefixed = embed_workspace_new(mprefixed); - if (!a || !b || !c || !wf32 || !wbf16 || !wprefixed) { + if (!a || !b || !c || !e || !wf32 || !wbf16 || !wf16 || !wprefixed) { fprintf(stderr, "allocation failed\n"); return 1; } if (embed_model_encode_into(mf32, wf32, ids, 4, a) != 0 || embed_model_encode_into(mbf16, wbf16, ids, 4, b) != 0 || + embed_model_encode_into(mf16, wf16, ids, 4, e) != 0 || embed_model_encode_into(mprefixed, wprefixed, ids, 4, c) != 0) { fprintf(stderr, "embedding failed\n"); return 1; } float max_diff = 0.0f; + float f16_diff = 0.0f; float prefix_diff = 0.0f; float norm = 0.0f; for (int i = 0; i < dim; i++) { - if (!isfinite(a[i]) || !isfinite(b[i]) || !isfinite(c[i])) { + if (!isfinite(a[i]) || !isfinite(b[i]) || !isfinite(e[i]) || !isfinite(c[i])) { fprintf(stderr, "non-finite embedding value\n"); return 1; } float d = fabsf(a[i] - b[i]); if (d > max_diff) max_diff = d; + d = fabsf(a[i] - e[i]); + if (d > f16_diff) + f16_diff = d; d = fabsf(a[i] - c[i]); if (d > prefix_diff) prefix_diff = d; @@ -77,10 +88,11 @@ int main(void) { } norm = sqrtf(norm); /* Embeddings are intentionally unnormalized; require a sane nonzero - * magnitude and tight F32-vs-BF16 parity, not unit norm. */ - if (norm <= 1e-6f || max_diff > 2e-5f || prefix_diff != 0.0f) { - fprintf(stderr, "bad parity: norm=%.9g max_diff=%.9g prefix_diff=%.9g\n", norm, max_diff, - prefix_diff); + * magnitude and tight F32-vs-BF16/F16 parity, not unit norm. F16 carries + * more mantissa bits than BF16, so the same 2e-5 bound holds. */ + if (norm <= 1e-6f || max_diff > 2e-5f || f16_diff > 2e-5f || prefix_diff != 0.0f) { + fprintf(stderr, "bad parity: norm=%.9g max_diff=%.9g f16_diff=%.9g prefix_diff=%.9g\n", + norm, max_diff, f16_diff, prefix_diff); return 1; } @@ -91,36 +103,43 @@ int main(void) { for (int i = 0; i < LONG_N; i++) long_ids[i] = (i % 14) + 1; if (embed_model_encode_into(mf32, wf32, long_ids, LONG_N, a) != 0 || - embed_model_encode_into(mbf16, wbf16, long_ids, LONG_N, b) != 0) { + embed_model_encode_into(mbf16, wbf16, long_ids, LONG_N, b) != 0 || + embed_model_encode_into(mf16, wf16, long_ids, LONG_N, e) != 0) { fprintf(stderr, "long-sequence embedding failed\n"); return 1; } - float long_diff = 0.0f; + float long_diff = 0.0f, long_f16_diff = 0.0f; for (int i = 0; i < dim; i++) { - if (!isfinite(a[i]) || !isfinite(b[i])) { + if (!isfinite(a[i]) || !isfinite(b[i]) || !isfinite(e[i])) { fprintf(stderr, "non-finite long-sequence value\n"); return 1; } float d = fabsf(a[i] - b[i]); if (d > long_diff) long_diff = d; + d = fabsf(a[i] - e[i]); + if (d > long_f16_diff) + long_f16_diff = d; } - if (long_diff > 2e-5f) { - fprintf(stderr, "bad widen-path parity: max_diff=%.9g\n", long_diff); + if (long_diff > 2e-5f || long_f16_diff > 2e-5f) { + fprintf(stderr, "bad widen-path parity: bf16=%.9g f16=%.9g\n", long_diff, long_f16_diff); return 1; } - printf("ok: bf16 and model-prefix parity dim=%d max_abs_diff=%.9g norm=%.9g\n", dim, max_diff, - norm); + printf("ok: bf16/f16 and model-prefix parity dim=%d bf16_diff=%.9g f16_diff=%.9g norm=%.9g\n", + dim, max_diff, f16_diff, norm); embed_workspace_free(wf32); embed_workspace_free(wbf16); + embed_workspace_free(wf16); embed_workspace_free(wprefixed); embed_model_free(mf32); embed_model_free(mbf16); + embed_model_free(mf16); embed_model_free(mprefixed); free(a); free(b); free(c); + free(e); return 0; } |
