feat: add native TensorRT-LLM P2P example

[zhengluo-nv]

Summary

• Add one production-style TensorRT-LLM Kubernetes example using the native checkpoint_format="MX" loader through the standard trtllm-serve --config interface.
• Remove the legacy patched PRESHARDED integration, including publish_from_worker, MxLiveCheckpointLoader, the TRT-LLM patch bundle, and historical Kimi/Dynamo-specific manifests.
• Remove the temporary TRT-LLM CI image, source overlay, merge/patch scripts, launcher, and executable GitHub Actions build/test entries.
• Keep the TRT-LLM row in ci/TEST_PLAN.md, marked Not started until a qualified upstream image and a no-shards fallback-proof test are available.
• Retain only the ModelExpress transport helpers consumed by TensorRT-LLM's native MX checkpoint loader.

Why

The repository contained two incompatible TensorRT-LLM paths: a historical patch-based integration and a temporary CI overlay assembled from multiple upstream revisions. Neither is appropriate as the canonical production path.

This change makes the upstream-native loader the only supported architecture. ModelExpress no longer patches or overlays TensorRT-LLM. Operators must use a qualified TensorRT-LLM image that already contains the native MX checkpoint loader, with the ModelExpress client installed without replacing the image's CUDA, Torch, NIXL, UCX, or protobuf stack.

User impact

The TensorRT-LLM example now matches the normal deployment shape used by the vLLM and SGLang examples:

• one scalable Deployment rather than dedicated source and target workloads;
• a shared Hugging Face model cache for normal fallback;
• checkpoint_format: MX and mx_config supplied through a standard TRT-LLM configuration file;
• the first replica loads from storage and publishes weights, while later compatible replicas can receive them through NIXL P2P.

TRT-LLM CI is intentionally not enabled by this PR. Re-enabling it requires a qualified upstream image and a target without local weight shards so storage fallback cannot produce a false pass.

Validation

• Full ModelExpress Python client suite: 740 passed, 29 skipped
• Focused TensorRT-LLM cleanup and native-loader tests: 8 passed
• GitHub workflow YAML parsing
• TensorRT-LLM Kubernetes manifest parsing
• Legacy path/reference audit
• git diff --check

Summary by CodeRabbit

• New Features

  • Added a native TensorRT-LLM P2P weight-transfer deployment using checkpoint_format="MX".
  • Added Kubernetes configuration for scalable single-node TensorRT-LLM replicas.
  • Added guidance for supported models, deployment, scaling, and verification.

• Documentation

  • Updated setup guides, integration references, and CI coverage to reflect native MX support.
  • Clarified vLLM and TensorRT-LLM deployment instructions.

• Chores

  • Removed legacy TensorRT-LLM preshared examples, patch tooling, and related deployment configurations.
  • Added automated validation for the new deployment and documentation.

[coderabbitai]

Walkthrough

Changes

Native MX TensorRT-LLM P2P

Layer / File(s)	Summary
Native MX runtime integration modelexpress_client/python/..., modelexpress_client/python/README.md	TensorRT-LLM documentation and transfer code now use native checkpoint_format="MX" loading and remove the custom PRESHARDED loader integration.
Native MX Kubernetes deployment examples/p2p_transfer_k8s/client/...	Adds the TensorRT-LLM ConfigMap, Service, Deployment, deployment guide, and tests for the single-node-per-replica MX path.
Documentation and contributor alignment README.md, CONTRIBUTING.md, examples/model_streamer_k8s/README.md	Updates feature descriptions, build instructions, and example links for the native MX workflow.
CI matrix and test-plan alignment .github/workflows/..., ci/...	Removes TensorRT-LLM CI matrix entries and updates artifact expectations, coverage status, and marker documentation.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Poem

I’m a rabbit with MX in my hat,
Hopping past old loaders—imagine that!
ConfigMaps bloom, replicas take flight,
CI trims its path through the night.
Native weights now flow swift and bright.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 16.67% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly and concisely summarizes the main change: adding a native TensorRT-LLM P2P example.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

modelexpress_client/python/tests/test_trtllm_native_example.py (1)

46-62: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Include the changed client README in the legacy-reference scan.

examples/p2p_transfer_k8s/client/README.md is a changed user-facing document, but it is absent from documentation. A stale llama-p2p.yaml or PRESHARDED reference there would pass this test. Add that file and keep the list aligned with the migration documentation.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@modelexpress_client/python/tests/test_trtllm_native_example.py` around lines
46 - 62, Update test_documentation_points_to_the_native_trtllm_example by adding
examples/p2p_transfer_k8s/client/README.md to the documentation list, ensuring
it is covered by the existing legacy-reference assertions and remains aligned
with the migration documentation.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@modelexpress_client/python/tests/test_trtllm_native_example.py`:
- Around line 46-62: Update
test_documentation_points_to_the_native_trtllm_example by adding
examples/p2p_transfer_k8s/client/README.md to the documentation list, ensuring
it is covered by the existing legacy-reference assertions and remains aligned
with the migration documentation.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: ca682b70-02a7-4f68-a034-df93a8530396

📥 Commits

Reviewing files that changed from the base of the PR and between f463897 and 6c17e00.

📒 Files selected for processing (26)

• .github/workflows/modelexpress-ci-tests.yml
• CONTRIBUTING.md
• README.md
• ci/TEST_PLAN.md
• ci/k8s/client/sgl/manifest-azure.yaml
• ci/k8s/client/test_p2p_k8s.py
• ci/k8s/client/trt-llm/Dockerfile
• ci/k8s/client/trt-llm/manifest-azure.yaml
• ci/k8s/client/trt-llm/trtllm_p2p_launcher.py
• examples/model_streamer_k8s/README.md
• examples/p2p_transfer_k8s/README.md
• examples/p2p_transfer_k8s/client/README.md
• examples/p2p_transfer_k8s/client/trtllm/Dockerfile.ph3-gcp-gb200
• examples/p2p_transfer_k8s/client/trtllm/README.md
• examples/p2p_transfer_k8s/client/trtllm/kimi-disagg-mx-tp8-dgd.yaml
• examples/p2p_transfer_k8s/client/trtllm/kimi-source-decode-dgd.yaml
• examples/p2p_transfer_k8s/client/trtllm/mx-infra-decode.yaml
• examples/p2p_transfer_k8s/client/trtllm/trtllm-single-node-p2p.yaml
• modelexpress_client/python/README.md
• modelexpress_client/python/modelexpress/envs.py
• modelexpress_client/python/modelexpress/trtllm_live_transfer.py
• modelexpress_client/python/tests/test_trtllm_native_example.py
• trtllm_patches/v1.3.0rc5/README.md
• trtllm_patches/v1.3.0rc5/apply_patches.py
• trtllm_patches/v1.3.0rc5/patch_model_loader.py
• trtllm_patches/v1.3.0rc5/patch_tp_allgather.py

💤 Files with no reviewable changes (13)

• ci/k8s/client/trt-llm/manifest-azure.yaml
• examples/p2p_transfer_k8s/client/trtllm/kimi-disagg-mx-tp8-dgd.yaml
• examples/p2p_transfer_k8s/client/trtllm/mx-infra-decode.yaml
• modelexpress_client/python/modelexpress/envs.py
• trtllm_patches/v1.3.0rc5/patch_tp_allgather.py
• examples/p2p_transfer_k8s/client/trtllm/Dockerfile.ph3-gcp-gb200
• trtllm_patches/v1.3.0rc5/README.md
• ci/k8s/client/trt-llm/trtllm_p2p_launcher.py
• ci/k8s/client/trt-llm/Dockerfile
• trtllm_patches/v1.3.0rc5/patch_model_loader.py
• trtllm_patches/v1.3.0rc5/apply_patches.py
• examples/p2p_transfer_k8s/client/trtllm/kimi-source-decode-dgd.yaml
• .github/workflows/modelexpress-ci-tests.yml