feat(vllm_dissag): Kimi-K3-MXFP4 2P/2D disagg framework integration

[MIR-AMD]

Summary

• Adds Kimi-K3-MXFP4 (MI300X gfx942 MXFP4 MoE, 2.8T params, 896 experts) to the vllm_dissag framework as a first-class disaggregated model
• Config-only integration: 10 files edited in scripts/vllm_dissag/, no standalone recipe folder
• Topology: 2P/2D (2 prefill + 2 decode nodes), TP2×DP8 per pool, MoRI-EP all2all, MoRIIO KV connector

Changes

• models.yaml: K3 recipe env anchor + model entry (20 env vars, prefill.dp/decode.dp flags)
• run_xPyD_models.slurm: add to VALID/MORI_EP_VALID/WIDE_EP_ONLY allowlists + 2P/2D validation + JIT cache split
• connectors/moriio.sh: _moriio_is_kimik3 gating (pod_hosts KV config, TP2×DP8, api-server-count, headless KV)
• vllm_disagg.sh: PREFILL/DECODE_POD_HOSTS export + K3 DP math + _model_config_to_array helper
• tests/argv_assert.sh: K3 offline assertions (proves K3 works AND non-K3 unchanged)

Regression safety

All K3 changes are guarded by _moriio_is_kimik3() / MODEL_NAME=="Kimi-K3-MXFP4". Three shared improvements (cache dir default, eval→_model_config_to_array, MAX_NUM_BATCHED_TOKENS flag) are model-agnostic and no-op for non-K3.

Offline argv_assert.sh validates: 37 passed, 0 failed (including existing DeepSeek/Llama assertions).

Image

Image build is documented in README.MD. Operators build separately from public repos:

docker build -f <path>/Dockerfile.kimik3_disagg \
  --build-arg VLLM_REF=kimi-k3-wideep-disagg-fullsource-v3 \
  -t kimik3-wideep-disagg:latest .

Related

• Overlaps with PR vllm_dissag: GLM-5.1-FP8 (MLA+DSA) MoRI-EP WideEP disaggregated enablement #206 (GLM) in models.yaml/moriio.sh/slurm/tests — additive, resolvable (K3 first, GLM rebases)
• Supersedes the K3 portion of draft PR Kimi-K3 MI300X (gfx942) 2P/2D disagg — lean recipe (config-only, fixes baked in image) #236 (which bundled 33 files including colocated recipes)

Test plan

• bash -n on all .sh files
• YAML lint on models.yaml
• argv_assert.sh offline tests (37/37 pass)
• Residue grep (zero internal paths)
• Scope gate (only scripts/vllm_dissag/* files)
• Optional: 4-node live smoke with NIAH

🤖 Generated with Claude Code