diff options
| author | Onur Gündüz <ogunduz@gmail.com> | 2026-06-10 07:30:34 +0200 |
|---|---|---|
| committer | Onur Gündüz <ogunduz@gmail.com> | 2026-06-14 08:27:03 +0200 |
| commit | c14151290df20e1a2ebc9fb06a77b075ca4ad959 (patch) | |
| tree | d4b53dd131f0e5f5f8f5dd60727cc13a76c42534 /pplx_server.c | |
| parent | dcc54e7a267f88ed456209c63f58c7130274e8da (diff) | |
| download | ffwd-c14151290df20e1a2ebc9fb06a77b075ca4ad959.tar.gz ffwd-c14151290df20e1a2ebc9fb06a77b075ca4ad959.zip | |
cli: rename flags, add --stream, split serving into pplx-embed-server
Diffstat (limited to 'pplx_server.c')
| -rw-r--r-- | pplx_server.c | 23 |
1 files changed, 10 insertions, 13 deletions
diff --git a/pplx_server.c b/pplx_server.c index 6e7bb22..f8ad5e1 100644 --- a/pplx_server.c +++ b/pplx_server.c @@ -2266,29 +2266,26 @@ static int mlx_memory_preflight(const pplx_server_config_t *cfg) { "skipping MLX memory preflight"); return 0; } - uint64_t budget = - physical / 100 * PPLX_MLX_MEMORY_BUDGET_PERCENT + - physical % 100 * PPLX_MLX_MEMORY_BUDGET_PERCENT / 100; + double utilization = cfg->memory_utilization > 0.0 + ? cfg->memory_utilization + : PPLX_MLX_MEMORY_BUDGET_PERCENT / 100.0; + uint64_t budget = (uint64_t)((double)physical * utilization); const double gib = 1024.0 * 1024.0 * 1024.0; server_log("pplx-serve: MLX memory preflight: %.1f GiB tensors, " - "%.1f GiB estimated resident, %.1f GiB safety budget", + "%.1f GiB estimated resident, %.1f GiB budget " + "(%.2f of physical memory)", (double)payload / gib, (double)estimated / gib, - (double)budget / gib); + (double)budget / gib, utilization); if (cfg->mlx_quantize_bits) { server_log("pplx-serve: MLX %d-bit quantization enabled; preflight " "uses source payload as a conservative peak estimate", cfg->mlx_quantize_bits); } if (estimated <= budget) return 0; - if (cfg->allow_memory_overcommit) { - server_log("pplx-serve: warning: MLX memory estimate exceeds safety " - "budget; continuing due to --allow-memory-overcommit"); - return 0; - } - server_log("pplx-serve: refusing MLX model set above host-memory safety " + server_log("pplx-serve: refusing MLX model set above the host-memory " "budget"); - server_log("pplx-serve: use BF16 artifacts, load fewer models, or pass " - "--allow-memory-overcommit"); + server_log("pplx-serve: use BF16 artifacts, load fewer models, or raise " + "--memory-utilization (values above 1.0 overcommit)"); return -1; } |
