diff --git a/circuitforge_core/resources/profiles/public/cpu-16gb.yaml b/circuitforge_core/resources/profiles/public/cpu-16gb.yaml new file mode 100644 index 0000000..75a22e8 --- /dev/null +++ b/circuitforge_core/resources/profiles/public/cpu-16gb.yaml @@ -0,0 +1,33 @@ +schema_version: 1 +name: cpu-16gb +eviction_timeout_s: 30.0 +services: + ollama: + max_mb: 0 + priority: 1 + cf-stt: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 1 + backend: moonshine + cf-tts: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 1 + cf-embed: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 2 + always_on: true + cf-classify: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 2 + always_on: true +model_size_hints: + llm_max_params: 3b-q4 + image_gen_max: none diff --git a/circuitforge_core/resources/profiles/public/cpu-32gb.yaml b/circuitforge_core/resources/profiles/public/cpu-32gb.yaml new file mode 100644 index 0000000..fb7c976 --- /dev/null +++ b/circuitforge_core/resources/profiles/public/cpu-32gb.yaml @@ -0,0 +1,33 @@ +schema_version: 1 +name: cpu-32gb +eviction_timeout_s: 30.0 +services: + ollama: + max_mb: 0 + priority: 1 + cf-stt: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 2 + backend: faster-whisper + cf-tts: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 2 + cf-embed: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 4 + always_on: true + cf-classify: + max_mb: 0 + priority: 2 + shared: true + max_concurrent: 4 + always_on: true +model_size_hints: + llm_max_params: 7b-q4 + image_gen_max: none diff --git a/circuitforge_core/resources/profiles/public/single-gpu-16gb.yaml b/circuitforge_core/resources/profiles/public/single-gpu-16gb.yaml new file mode 100644 index 0000000..7ad59f9 --- /dev/null +++ b/circuitforge_core/resources/profiles/public/single-gpu-16gb.yaml @@ -0,0 +1,45 @@ +schema_version: 1 +name: single-gpu-16gb +vram_total_mb: 16384 +eviction_timeout_s: 10.0 +services: + vllm: + max_mb: 12288 + priority: 1 + ollama: + max_mb: 12288 + priority: 1 + cf-vision: + max_mb: 3072 + priority: 2 + shared: true + max_concurrent: 4 + cf-stt: + max_mb: 1200 + priority: 2 + shared: true + max_concurrent: 3 + backend: parakeet-tdt + cf-tts: + max_mb: 1024 + priority: 2 + shared: true + max_concurrent: 3 + cf-embed: + max_mb: 512 + priority: 2 + shared: true + max_concurrent: 6 + always_on: true + cf-classify: + max_mb: 512 + priority: 2 + shared: true + max_concurrent: 6 + always_on: true + comfyui: + max_mb: 14336 + priority: 4 +model_size_hints: + llm_max_params: 34b + image_gen_max: flux-dev-fp8 diff --git a/circuitforge_core/resources/profiles/public/single-gpu-24gb.yaml b/circuitforge_core/resources/profiles/public/single-gpu-24gb.yaml new file mode 100644 index 0000000..4f98eb8 --- /dev/null +++ b/circuitforge_core/resources/profiles/public/single-gpu-24gb.yaml @@ -0,0 +1,45 @@ +schema_version: 1 +name: single-gpu-24gb +vram_total_mb: 24576 +eviction_timeout_s: 10.0 +services: + vllm: + max_mb: 20480 + priority: 1 + ollama: + max_mb: 18432 + priority: 1 + cf-vision: + max_mb: 4096 + priority: 2 + shared: true + max_concurrent: 6 + cf-stt: + max_mb: 1200 + priority: 2 + shared: true + max_concurrent: 4 + backend: parakeet-tdt + cf-tts: + max_mb: 1024 + priority: 2 + shared: true + max_concurrent: 4 + cf-embed: + max_mb: 512 + priority: 2 + shared: true + max_concurrent: 8 + always_on: true + cf-classify: + max_mb: 512 + priority: 2 + shared: true + max_concurrent: 8 + always_on: true + comfyui: + max_mb: 20480 + priority: 4 +model_size_hints: + llm_max_params: 70b + image_gen_max: flux-dev-fp16 diff --git a/circuitforge_core/resources/profiles/public/single-gpu-4gb.yaml b/circuitforge_core/resources/profiles/public/single-gpu-4gb.yaml new file mode 100644 index 0000000..b7cb24e --- /dev/null +++ b/circuitforge_core/resources/profiles/public/single-gpu-4gb.yaml @@ -0,0 +1,30 @@ +schema_version: 1 +name: single-gpu-4gb +vram_total_mb: 4096 +eviction_timeout_s: 15.0 +services: + ollama: + max_mb: 3072 + priority: 1 + cf-vision: + max_mb: 1024 + priority: 2 + shared: true + max_concurrent: 1 + cf-stt: + max_mb: 600 + priority: 2 + shared: true + max_concurrent: 1 + backend: faster-whisper + cf-tts: + max_mb: 512 + priority: 2 + shared: true + max_concurrent: 1 + comfyui: + max_mb: 3584 + priority: 4 +model_size_hints: + llm_max_params: 3b + image_gen_max: sd15-fp8