Upload README.md

README.md

@@ -0,0 +1,264 @@
+---
+title: Inference Capacity Planner
+emoji: ⚡
+colorFrom: purple
+colorTo: yellow
+sdk: static
+app_file: index.html
+pinned: false
+license: mit
+short_description: Plan LLM inference capacity, VRAM, throughput and cost.
+---
+
+# ⚡ Inference Capacity Planner
+
+**Plan LLM serving capacity before you deploy.**
+
+A lightweight, local-first tool for estimating **model memory, sustainable throughput, required replicas, and serving cost** for LLM inference workloads.
+
+Built by the independent Hugging Face organization **Inferencing**.
+
+---
+
+## What it helps you estimate
+
+Use the planner to get a quick first-pass view of:
+
+- model weight memory
+- approximate runtime memory
+- tokens per request
+- sustainable requests per second
+- required replicas
+- total GPU count
+- approximate generation time
+- daily token volume
+- hourly infrastructure cost
+- daily infrastructure cost
+- monthly infrastructure cost
+- approximate cost per 1M tokens
+- capacity headroom
+
+---
+
+## Configure your workload
+
+The Space lets you adjust:
+
+### Model
+
+- parameter count
+- weight precision
+- runtime memory multiplier
+
+### Traffic
+
+- input tokens per request
+- output tokens per request
+- target requests per second
+
+### Serving performance
+
+- aggregate tokens per second
+- target utilization
+- time to first token
+
+### Infrastructure
+
+- GPUs per replica
+- GPU price per hour
+
+---
+
+## Quick presets
+
+The interface includes starting profiles for:
+
+- **~8B models**
+- **~30B models**
+- **~70B models**
+
+These presets are intentionally generic.
+
+They are not tied to a specific model, GPU, inference engine, or provider.
+
+---
+
+## Local-first
+
+> 🔒 **All calculations happen directly in your browser.**
+
+The Space does not upload or store:
+
+- workload assumptions
+- model settings
+- token volumes
+- infrastructure costs
+- pricing inputs
+
+No backend or external API is required.
+
+---
+
+## Core formulas
+
+### Model weight memory
+
+```text
+weight_memory_GB
+=
+parameters_billions × bytes_per_weight
+```
+
+### Runtime memory estimate
+
+```text
+runtime_memory_GB
+=
+weight_memory_GB × runtime_overhead
+```
+
+### Sustainable requests per second
+
+```text
+tokens_per_second × utilization
+────────────────────────────────
+input_tokens + output_tokens
+```
+
+### Required replicas
+
+```text
+target_requests_per_second
+──────────────────────────
+sustainable_RPS_per_replica
+```
+
+The result is rounded up.
+
+### Estimated infrastructure cost
+
+```text
+hourly_cost
+=
+replicas × GPUs_per_replica × GPU_hourly_price
+```
+
+---
+
+# Why this Space exists
+
+Model quality is only one part of production AI.
+
+A useful inference system must also balance:
+
+```text
+QUALITY
+   ▲
+   │
+   │
+   └────────────► SPEED
+        \
+         \
+          COST
+```
+
+The largest model is not always the best deployment choice.
+
+Inference planning is often about finding the right operating point between:
+
+**quality · latency · throughput · memory · reliability · cost**
+
+---
+
+## Useful for
+
+This Space can help with:
+
+- early architecture discussions
+- deployment planning
+- cost estimation
+- comparing model sizes
+- comparing quantization profiles
+- capacity sanity checks
+- estimating scale requirements
+- teaching inference economics
+- planning benchmarks
+- prototyping
+
+---
+
+## Important limitations
+
+This is a **planning calculator**, not a hardware benchmark.
+
+Real-world inference performance can vary substantially depending on:
+
+- model architecture
+- exact quantization format
+- GPU model
+- inference runtime
+- context length
+- KV cache usage
+- batching
+- concurrency
+- tensor parallelism
+- memory bandwidth
+- speculative decoding
+- CUDA configuration
+- framework overhead
+- prompt distribution
+- output length
+
+For production decisions, benchmark your **exact model + runtime + hardware + workload**.
+
+---
+
+## Future directions
+
+Possible extensions include:
+
+- GPU comparison
+- multi-GPU planning
+- KV-cache estimation
+- concurrency modeling
+- provider cost comparison
+- batch-size simulation
+- quantization comparison
+- latency targets
+- model routing scenarios
+- edge inference profiles
+- downloadable deployment reports
+
+---
+
+## About Inferencing
+
+**Inferencing** explores the infrastructure layer that turns AI models into usable systems.
+
+Topics include:
+
+- model serving
+- latency
+- throughput
+- batching
+- caching
+- quantization
+- routing
+- edge inference
+- serving economics
+- observability
+- deployment efficiency
+
+---
+
+## Independent project
+
+**Inferencing is an independent Hugging Face community organization.**
+
+It is not an official Hugging Face organization, cloud provider, model provider, benchmark authority, or infrastructure vendor.
+
+---
+
+# Inferencing
+
+### **Serve faster. Route smarter. Scale intelligence.**