feat: add P/D disaggregation output to llm-d generator

[amito]

Description

Adds prefill/decode (P/D) disaggregation as an output option to the llm-d deployment generator. When P/D is enabled, the generator produces separate Kustomize patches for prefill and decode deployments instead of the single patch-vllm.yaml, allowing users to configure independent replica counts and vLLM arguments for each role.

When pd_enabled=False (the default), output is identical to the current behavior introduced in PR #274. No breaking changes.

What gets generated when P/D is enabled

• patch-prefill.yaml - Prefill deployment with --kv-role=kv_producer, --enable-chunked-prefill, and configurable prefill_replicas
• patch-decode.yaml - Decode deployment with --kv-role=kv_consumer and configurable decode_replicas
• kustomization.yaml - Conditionally references prefill + decode patches instead of the single vLLM patch
• values.yaml - EPP Helm values (unchanged)

KV transfer uses nixlv2 (NixlConnector) with separate --kv-connector and --kv-role args, matching the llm-d P/D well-lit path.

Template consolidation

All three patch templates (patch-vllm, patch-prefill, patch-decode) are consolidated into a single patch-modelserver.yaml.j2 rendered with different context variables (deployment_name, replica_count, extra_args). This avoids maintaining three near-identical templates that would drift independently. Role-specific differences (e.g., kv-connector args, chunked-prefill) are passed as extra_args from the generator code. If the prefill and decode roles diverge significantly in the future (different sidecars, volumes, probes), splitting back into separate templates is a one-step refactor.

Changes

• src/planner/configuration/templates/llmd/patch-modelserver.yaml.j2 - Unified template for all model-server patches (replaces patch-vllm, patch-prefill, patch-decode).
• src/planner/configuration/llmd_generator.py - generate_all() accepts pd_enabled, prefill_replicas, and decode_replicas parameters. Renders the unified template with role-specific context. Validates replica counts >= 1 at the generator level (raises ValueError).
• src/planner/configuration/templates/llmd/kustomization.yaml.j2 - Conditional patch references based on pd_enabled.
• src/planner/api/routes/configuration.py - Added pd_enabled, prefill_replicas, and decode_replicas fields to DeploymentRequest with Field(ge=1, le=32) validation on replica counts.
• ui/components/deployment.py - Added "Enable P/D Disaggregation" checkbox with prefill/decode replica inputs (shown when enabled, llm-d stack only). File display section handles patch_prefill/patch_decode keys.
• tests/unit/test_llmd_generator.py - Added TestPDDisaggregation (8 tests), TestDeployAPINewParams (4 tests). 31 tests total.

Note on deployment_name: The non-PD patch uses metadata.name: decode (not vllm) because it is a strategic merge patch targeting the upstream llm-d base Deployment/decode. A dedicated test locks this down.

How Has This Been Tested?

12 new unit tests added, full suite passes (31 tests in test_llmd_generator.py):

cd src && uv run pytest ../tests/unit/test_llmd_generator.py -v

Tests cover:

• test_pd_disabled_produces_single_patch - Default behavior unchanged: patch_vllm present, no patch_prefill/patch_decode.
• test_pd_disabled_patch_targets_decode_deployment - Non-PD patch uses metadata.name: decode to match the llm-d base Deployment.
• test_pd_enabled_produces_prefill_and_decode_patches - P/D mode produces patch_prefill and patch_decode, no patch_vllm.
• test_pd_prefill_patch_has_correct_replicas - Prefill patch uses prefill_replicas value (non-default) and has metadata.name: prefill.
• test_pd_decode_patch_has_correct_replicas - Decode patch uses decode_replicas value and has metadata.name: decode.
• test_pd_kustomization_references_both_patches - Kustomization patches list contains patch-prefill.yaml and patch-decode.yaml, not patch-vllm.yaml.
• test_pd_all_outputs_valid_yaml - All generated contents parse as valid YAML.
• test_rejects_zero_replicas (parametrized) - Generator raises ValueError when prefill_replicas or decode_replicas < 1.
• test_deploy_llmd_with_pd_enabled - API integration test: POST with pd_enabled: true returns 200 with patch_prefill and patch_decode.
• test_deploy_rejects_zero_replicas (parametrized) - API returns 422 when prefill_replicas or decode_replicas is 0.
• test_deploy_rejects_replicas_above_max - API returns 422 when replica count exceeds 32.

UI manually tested: selecting llm-d stack and enabling the P/D checkbox shows prefill/decode replica inputs; generated files display correctly with the new file labels.

Merge criteria:

• The commits are squashed in a cohesive manner and have meaningful messages.
• Testing instructions have been added in the PR body (for PRs involving changes that are not immediately obvious).
• The developer has manually tested the changes and verified that the changes work

[jgchn]

Could be a future improvement, but the "Deploy to Kubernetes" path doesn't forward the new params like pd_enabled,prefill_replicas and decode_replicas.

[amito]

Fixed, thanks.

[jgchn]

I think we need to clear the stale YAMLs after user checks this box?

[amito]

Agreed, fixing this, thanks.

[jgchn]

Not a bug but the generation checks for two different ranges. The Pydantic model in configuration.py checks for Field(1, ge=1, le=32) but here, the generation checks for just <1. Perhaps we should remain consistent?

[amito]

I agree. Fixing this for consistency. Thanks