diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-27 00:58:05 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-27 01:03:41 +0200 |
| commit | af55803d0dff9eb692c760c8d486e736c945061d (patch) | |
| tree | 0875a1112bf98304e445e0544c4309670d6f4769 /libffwd/mlx.c | |
| parent | 0f861b7f3098750ddee4f6570f3be228d1da5320 (diff) | |
| download | ffwd-af55803d0dff9eb692c760c8d486e736c945061d.tar.gz ffwd-af55803d0dff9eb692c760c8d486e736c945061d.zip | |
chore: clean up comment formatting and wording
Diffstat (limited to 'libffwd/mlx.c')
| -rw-r--r-- | libffwd/mlx.c | 57 |
1 files changed, 25 insertions, 32 deletions
diff --git a/libffwd/mlx.c b/libffwd/mlx.c index 0b6d8b7..d8f6d1e 100644 --- a/libffwd/mlx.c +++ b/libffwd/mlx.c @@ -13,9 +13,7 @@ #include <limits.h> #include <float.h> -/* ======================================================================== - * Per-layer MLX weight arrays - * ======================================================================== */ +/* Per-layer MLX weight arrays */ typedef struct { mlx_array w; @@ -33,9 +31,9 @@ typedef struct { mlx_array q_norm, k_norm; mlx_array input_norm, post_attn_norm; mlx_linear_t gate_proj, up_proj, down_proj; - /* BERT family (unset for Qwen3; see internal.h). wo/up_proj/down_proj - * and input_norm/post_attn_norm are reused as the BERT attention output, - * FFN matrices, and the two block LayerNorm weights. */ + /* BERT family only; unset for Qwen3. See internal.h. + * Reuses wo/up_proj/down_proj for BERT attention output and FFN matrices. + * Reuses input_norm/post_attn_norm for the two BERT block LayerNorm weights. */ mlx_array o_bias, attn_ln_bias, ffn_inter_bias, ffn_out_bias, ffn_ln_bias; } mlx_layer_t; @@ -44,7 +42,7 @@ struct ffwd_mlx_ctx { mlx_array embed_tokens; mlx_layer_t *layers; /* heap-allocated [n_layers] */ mlx_array norm; - /* BERT family (unset for Qwen3): learned positions, token-type, embed LN. */ + /* BERT family only; unset for Qwen3: learned positions, token-type, embed LN. */ mlx_array position_embeddings, token_type_embeddings, ffwd_ln_w, ffwd_ln_b; mlx_stream stream; multi_safetensors_t *ms; @@ -69,9 +67,7 @@ struct ffwd_mlx_late_vectors { int dim; }; -/* ======================================================================== - * Helpers - * ======================================================================== */ +/* Helpers */ static int arr_ok(mlx_array a) { return a.ctx != NULL; } @@ -158,10 +154,11 @@ static mlx_array load_tensor(multi_safetensors_t *ms, const char *name, const in if (!mlx_tensor_has_supported_shape(sf, t, name, shape, ndim)) return (mlx_array){0}; if (t->dtype == DTYPE_F16) { - /* No F16 forward path on MLX (the matmul fast path keys on BF16). Widen - * to F32 on the host at load so the whole forward runs in F32, mirroring - * the CPU loader. mlx_array_new_data copies its buffer, so the temporary - * is freed immediately. Public F16 snapshots load through here. */ + /* No F16 forward path on MLX; the matmul fast path uses BF16. + * Widen to F32 on the host at load time, so the full forward pass runs in F32, + * matching the CPU loader. + * mlx_array_new_data copies its buffer, so the temporary is freed immediately. + * Public F16 snapshots load through this path. */ int64_t n = safetensor_numel(t); float *tmp = (float *)malloc((size_t)n * sizeof(float)); if (!tmp) @@ -341,9 +338,7 @@ static void upcast_f32_inplace(mlx_array *a, mlx_stream s) { *a = f; } -/* ======================================================================== - * Load - * ======================================================================== */ +/* Load */ static int mlx_model_dir_has_file(const char *model_dir, const char *rel) { char path[1024]; @@ -394,9 +389,9 @@ static int prepare_layer_transposes(mlx_layer_t *l, mlx_stream s) { prepare_linear_transpose(&l->down_proj, s); } -/* BERT-family MLX loader, mirroring load_bert_weights on the CPU side. The - * shared post-load below (stream, transposes, optional BF16 upcast) runs after - * either family fills ctx->layers. */ +/* BERT-family MLX loader, matching CPU load_bert_weights. + * The shared post-load step below runs after either family fills ctx->layers: + * stream setup, transposes, and optional BF16 upcast. */ static int load_mlx_bert_weights(ffwd_mlx_ctx_t *ctx, int layer_start, int layer_end) { const ffwd_config_t *c = &ctx->config; int h = c->hidden_size, inter = c->intermediate_size; @@ -625,9 +620,11 @@ static ffwd_mlx_ctx_t *mlx_load_range_ex(const char *model_dir, loaded: ctx->stream = mlx_default_gpu_stream_new(); ctx->weight_dtype = mlx_array_dtype(ctx->layers[layer_start].wq.w); - /* Non-quantized BF16: run RMSNorm and the residual stream in F32. Upcast the - * tiny norm vectors so they match the F32 activations in mlx_forward_layers; - * the large matmul weights stay BF16. Q8 keeps its existing BF16 norms. */ + /* Non-quantized BF16: + * run RMSNorm and the residual stream in F32. + * Upcast the small norm vectors to match the F32 activations in + * mlx_forward_layers; keep the large matmul weights in BF16. + * Q8 keeps its existing BF16 norms. */ if (ctx->weight_dtype == MLX_BFLOAT16 && !ctx->quantize_bits) { for (int i = layer_start; i < layer_end; i++) { mlx_layer_t *l = &ctx->layers[i]; @@ -647,7 +644,7 @@ loaded: upcast_f32_inplace(&l->ffn_ln_bias, ctx->stream); } upcast_f32_inplace(&ctx->norm, ctx->stream); - /* BERT embedding stage (no-ops for Qwen): learned position and token-type + /* BERT embedding stage; no-ops for Qwen: learned position and token-type * embeddings and the embedding LayerNorm all feed the F32 residual. */ upcast_f32_inplace(&ctx->position_embeddings, ctx->stream); upcast_f32_inplace(&ctx->token_type_embeddings, ctx->stream); @@ -704,9 +701,7 @@ ffwd_mlx_ctx_t *ffwd_mlx_load_with_options(const char *model_dir, const ffwd_mlx return mlx_load_range_ex(model_dir, 0, -1, opts, 0); } -/* ======================================================================== - * Free - * ======================================================================== */ +/* Free */ static void free_mlx_layer(mlx_layer_t *l) { free_linear(&l->wq); @@ -827,9 +822,7 @@ const ffwd_config_t *ffwd_mlx_late_config(const ffwd_mlx_late_ctx_t *ctx) { int ffwd_mlx_late_token_dim(const ffwd_mlx_late_ctx_t *ctx) { return ctx ? ctx->token_dim : 0; } -/* ======================================================================== - * Forward pass - * ======================================================================== */ +/* Forward pass */ /* Exact erf GeLU: 0.5 * x * (1 + erf(x / sqrt(2))). Frees nothing it is given. */ static mlx_array mlx_gelu_erf(mlx_array x, mlx_stream S) { @@ -1569,7 +1562,7 @@ cleanup: /* Device-resident select/concat primitives. The server rerank path batches * candidates through ffwd_mlx_late_encode_docs_device, but the late-interaction - * verification harness (tests/integration/check_late_interaction.py) drives these directly + * verification harness (tests/check_late_interaction.py) drives these directly * to validate per-document gather and concat against the CPU MaxSim reference. */ ffwd_mlx_late_vectors_t *ffwd_mlx_late_vectors_concat(ffwd_mlx_late_ctx_t *ctx, const ffwd_mlx_late_vectors_t *const *items, @@ -2125,7 +2118,7 @@ static int ffwd_mlx_encode_batch_dense(ffwd_mlx_ctx_t *ctx, mlx_array_free(x_normed); } - /* 5. Eval and copy [B, hidden] to CPU. Perplexity embeddings are + /* 5. Eval and copy [B, hidden] to CPU. pplx-embed embeddings are * intentionally unnormalized; callers should use cosine similarity. */ mlx_array emb_f32 = mlx_array_new(); mlx_astype(&emb_f32, emb, MLX_FLOAT32, S); |
