feat: PD disaggregation uses client-provided config instead of DB recommendation (#1240)

Co-authored-by: cemeng <cemengzhang@gmail.com>

Author: Rader

common/types/model.go

@@ -282,6 +282,11 @@ type ModelRunReq struct {
 	// When true, the system checks the model metadata for PD recommendation,
 	// validates hardware resources, and splits resources between prefill and decode.
 	EnablePD bool `json:"enable_pd"`
+	// PD is the client-provided PD (Prefill-Decode) disaggregation configuration.
+	// When EnablePD is true, the client sends TP/DP/EP/PodsSize for prefill and decode
+	// roles. The server validates the config against available hardware resources
+	// instead of deriving it from PDRecommendation.
+	PD *PDConfig `json:"pd,omitempty"`
 	// OwnerNamespace is optional. If set, the inference is created under this namespace (user or org) for billing and listing; path {namespace} remains the model's owner.
 	OwnerNamespace string `json:"owner_namespace,omitempty"`
 }
@@ -312,6 +317,11 @@ type InstanceRunReq struct {
 	// When true, the system checks the model metadata for PD recommendation,
 	// validates hardware resources, and splits resources between prefill and decode.
 	EnablePD bool `json:"enable_pd"`
+	// PD is the client-provided PD (Prefill-Decode) disaggregation configuration.
+	// When EnablePD is true, the client sends TP/DP/EP/PodsSize for prefill and decode
+	// roles. The server validates the config against available hardware resources
+	// instead of deriving it from PDRecommendation.
+	PD *PDConfig `json:"pd,omitempty"`
 	// OwnerNamespace is optional. If set, the finetune is created under this namespace (user or org); path {namespace} remains the model's owner.
 	OwnerNamespace string `json:"owner_namespace,omitempty"`
 }

common/types/pd.go

@@ -190,17 +190,21 @@ type PDConfig struct {
 }
 
 // PDRoleRuntimeConfig holds the runtime parallelism and hardware resource
-// configuration for one PD role (prefill or decode). This is derived from
-// the PDRecommendation and the user's hardware input, split proportionally.
+// configuration for one PD role (prefill or decode). This is provided by the
+// client and validated/supplemented by the server.
 type PDRoleRuntimeConfig struct {
 	// TP is the tensor parallelism degree.
 	TP int `json:"tp"`
 	// EP is the expert parallelism degree (1 for dense models).
 	EP int `json:"ep"`
 	// DP is the data parallelism degree.
 	DP int `json:"dp"`
-	// TotalGPUs is the total GPUs for this role (TP * EP * DP).
+	// TotalGPUs is the total GPUs for this role (TP * DP). EP does not add extra GPUs.
 	TotalGPUs int `json:"total_gpus"`
+	// PodsSize is the number of pods per LWS group (maps to LWS spec.leaderWorkerTemplate.size).
+	// Each pod runs one vLLM/SGLang instance. GPUs per pod = TotalGPUs / PodsSize.
+	// When PodsSize is 0 or 1, all GPUs are in a single pod.
+	PodsSize int `json:"pods_size,omitempty"`
 	// Hardware is the hardware resource allocated to this role.
 	Hardware HardWare `json:"hardware"`
 }