pdd usage doc

docs/Prefill-decoding_Disaggregation.md

@@ -0,0 +1,25 @@
+# Prefill-decoding Disaggregation
+Existing large language model (LLM) serving systems that support prefill-decoding disaggregation typically compute the prefill and decoding phases on separate instances. For each request, following the prefill phase, the system migrates the generated key-value (KV) cache to the decoding instance and continues the computation. 
+
+
+Llumnix can manage the scheduling of runtime requests across instances. It can perceive the execution process of requests and the state of the KV cache. Llumnix provides a series of scheduling semantics to support the initial dispatch and runtime rescheduling (request migration across instances). 
+
+
+Therefore, Llumnix is inherently well-suited for the "special" cross-instance scheduling requirements in prefill-decoding disaggregation. Specifically, it limits the initial **dispatch only to the prefill instances** and the **migration between prefill and decoding instances**. 
+
+## Supported Features
+1. Requests can be **automatically migrated** from prefill instance to decoding instances.
+
+2. Users can specify the number of prefill instances.
+
+3. Llumnix supports both **one-to-many and many-to-one migrations** from prefill to decoding instances.
+
+4. Decoding instances can migrate requests among themselves based on different scheduling strategies (e.g. load-balance).
+
+
+## How to use
+Llumnix only uses two arguments to enable prefill-decoding disaggregation for simplicity. 
+- `--enable-pd-disagg True` is used to enable prefill-decoding disaggregation.
+- `--num-available-dispatch-instances` is used to configure the number of prefill instances. 
+
+Note that `num-available-dispatch-instances` < `initial_instance-num` especially when `--enable-scaling` is not set, as it determines the number of decoding instances.