fix(e2e): unbreak trtllm lanes broken by mcp 2.0.0a1 by slin1237 · Pull Request #1675 · smg-project/smg

slin1237 · 2026-06-11T19:47:44Z

Summary

Every trtllm e2e job has been failing at the model-download step (before any test runs) since the TensorRT-LLM 1.3.0rc18 bump (#1663):

File ".../e2e_test/infra/__init__.py", line 44, in <module>
    from .mock_mcp import ...
File ".../e2e_test/infra/mock_mcp_server.py", line 60, in <module>
    from mcp.server.fastmcp import FastMCP
ModuleNotFoundError: No module named 'mcp.server.fastmcp'

Root cause (with receipts from job 80871944044)

tensorrt-llm==1.3.0rc18 newly depends on mcp (unconstrained) — rc14 did not (Collecting mcp (from tensorrt-llm==1.3.0rc18)).
ci_install_trtllm.sh installs with pip install --pre (required for NVIDIA pre-release wheels) — which lets every transitive dep resolve to pre-releases → mcp-2.0.0a1 (note the neighbours: pydantic-2.14.0a1, lxml-7.0.0a2).
mcp 2.0 removed the legacy mcp.server.fastmcp module (verified by inspecting the wheel: zero fastmcp entries).
Our floor mcp>=1.0 treats the alpha as already satisfied (Requirement already satisfied: mcp>=1.0 ... (2.0.0a1)), so the e2e deps install keeps it.
e2e_test/infra/__init__.py eagerly imports the FastMCP-based mock — used only by the sglang-only responses/ lane — so even ci_download_model.sh (which just wants model specs) crashes.

Passing rc14 runs confirm the contrast: there mcp came only from e2e deps and resolved to stable mcp-1.27.2.

Fix

mcp>=1.2,<2 — the honest requirement (mcp.server.fastmcp was added in 1.2, removed in 2.0). Per PEP 440, <2 excludes pre-releases of 2.0 (verified with packaging: SpecifierSet(">=1.2,<2").contains("2.0.0a1", prereleases=True) == False), so pip install e2e_test/ now replaces the alpha with stable 1.x. TensorRT-LLM's own mcp requirement is unconstrained, so the downgrade leaves the environment consistent.
Lazy-load the mock MCP server via module __getattr__ in infra/__init__.py — model download and chat/router lanes no longer depend on the mcp SDK at all. The responses/ consumers (from infra import MockMcpServer, from infra.mock_mcp import mock_mcp_server) are unaffected.

Verification

mcp 2.0.0a1 wheel inspected: no mcp/server/fastmcp content; specifier semantics verified with packaging.
ruff check/format clean.
This PR's own e2e-1gpu-chat (trtllm) lane is the live regression test — it exercises the exact failing path (engine setup → e2e deps → model download → tests).

Summary by CodeRabbit

Chores
- Updated MCP dependency version constraints to ensure E2E test compatibility.
- Optimized test infrastructure module initialization with lazy loading.

Every trtllm e2e job dies at the model-download step since the TensorRT-LLM 1.3.0rc18 bump (#1663): rc18 newly depends on mcp (unconstrained), the trtllm install runs pip with --pre (required for NVIDIA pre-release wheels), and pip resolves mcp to 2.0.0a1 — which removed the legacy mcp.server.fastmcp module. Our e2e dependency floor mcp>=1.0 treats the alpha as already satisfied, and e2e_test/infra eagerly imports the FastMCP-based mock at package import, so even ci_download_model.sh crashes with ModuleNotFoundError before any test runs. rc14 never depended on mcp, which is why this only surfaced now. Two fixes, both on our side of the contract: - Declare the real requirement: mcp>=1.2,<2 (fastmcp was added in 1.2 and removed in 2.0). PEP 440 excludes 2.0 pre-releases from <2, so the e2e deps install now replaces the alpha with stable 1.x instead of keeping it. TensorRT-LLM's own mcp requirement is unconstrained, so the downgrade stays consistent. - Stop importing the mock MCP server eagerly in infra/__init__: only the sglang-only responses lane uses it, yet the eager import made model download and every chat/router lane depend on the mcp SDK. The symbols now resolve lazily via module __getattr__. Signed-off-by: Simo Lin <25425177+slin1237@users.noreply.github.com>

coderabbitai · 2026-06-11T19:48:00Z

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro

Run ID: 97a2f4f2-2584-48e7-998b-98e5bcc3b70c

📥 Commits

Reviewing files that changed from the base of the PR and between bb6373b and 7861ebd.

📒 Files selected for processing (2)

e2e_test/infra/__init__.py
e2e_test/pyproject.toml

📝 Walkthrough

Walkthrough

The pull request defers import of mock MCP server symbols in e2e_test.infra by replacing eager imports with a module-level __getattr__ lazy resolver, and pins the mcp dependency to versions 1.2–1.9 to ensure the required mcp.server.fastmcp module is available.

Changes

Lazy Mock MCP Resolution

Layer / File(s)	Summary
Lazy attribute resolution for mock MCP exports `e2e_test/infra/__init__.py`	Removed eager top-level import of `IMAGE_GENERATION_PNG_BASE64`, `MockMcpServer`, and `mock_mcp_server` from `.mock_mcp`, and added a module-level `__getattr__` function that imports `.mock_mcp` and returns the requested symbol only when one of the lazy exports is first accessed.
MCP version constraint `e2e_test/pyproject.toml`	Updated the mcp dependency from `>=1.0` to `>=1.2,<2` with an inline comment documenting that the version range is required because `infra.mock_mcp_server` depends on `mcp.server.fastmcp`, which was added in version 1.2 and removed in version 2.0.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Suggested labels

tests

Suggested reviewers

CatherineSue
key4ng
XinyueZhang369

Poem

🐰 Lazy imports hop by one by one,
Only loaded when the deed is done,
Version bounds keep fastmcp in sight,
No eager leaps—just-in-time feels right! ✨

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% 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 'fix(e2e): unbreak trtllm lanes broken by mcp 2.0.0a1' clearly describes the main fix: addressing broken e2e test lanes caused by mcp 2.0.0a1 incompatibility, which is the core issue resolved by the PR.
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.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

📝 Generate docstrings

Create stacked PR
Commit on current branch

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch fix/e2e-mcp-constraint

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

gemini-code-assist

Code Review

This pull request introduces lazy loading for mock MCP server symbols in e2e_test/infra/__init__.py to avoid importing the mcp SDK at import time, and restricts the mcp dependency version to mcp>=1.2,<2 in e2e_test/pyproject.toml. The reviewer suggested optimizing the lazy loading implementation by caching the resolved attributes in the module's globals to bypass subsequent __getattr__ lookups.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

gemini-code-assist · 2026-06-11T19:49:17Z

+def __getattr__(name: str) -> object:
+    if name in _LAZY_MOCK_MCP:
+        from . import mock_mcp
+
+        return getattr(mock_mcp, name)
+    raise AttributeError(f"module {__name__!r} has no attribute {name!r}")


To optimize module attribute lookups, we can cache the resolved attributes in the module's globals. This ensures that __getattr__ is only triggered once per attribute, and subsequent accesses will bypass __getattr__ entirely by retrieving the value directly from the module's namespace.

Suggested change

def __getattr__(name: str) -> object:

if name in _LAZY_MOCK_MCP:

from . import mock_mcp

return getattr(mock_mcp, name)

raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

def __getattr__(name: str) -> object:

if name in _LAZY_MOCK_MCP:

from . import mock_mcp

value = getattr(mock_mcp, name)

globals()[name] = value

return value

raise AttributeError(f"module {__name__!r} has no attribute {name!r}")

claude · 2026-06-11T19:50:05Z

👋 The PR description doesn't fully follow
PULL_REQUEST_TEMPLATE.md:

Missing header: ## Description
Missing header: ### Problem
Missing header: ### Solution
Missing header: ## Changes
Missing header: ## Test Plan

Please update the PR description so reviewers have the context they need.

claude

Clean fix — lazy-loading the mock MCP symbols via __getattr__ and pinning mcp>=1.2,<2 are both correct. No issues found.

slin1237 requested review from CatherineSue, XinyueZhang369 and key4ng as code owners June 11, 2026 19:47

github-actions Bot added dependencies Dependency updates tests Test changes labels Jun 11, 2026

gemini-code-assist Bot reviewed Jun 11, 2026

View reviewed changes

claude Bot approved these changes Jun 11, 2026

View reviewed changes

coderabbitai Bot approved these changes Jun 11, 2026

View reviewed changes

slin1237 merged commit 87d01f0 into main Jun 11, 2026
39 checks passed

slin1237 deleted the fix/e2e-mcp-constraint branch June 11, 2026 20:14

Navigation Menu

Repository navigation

fix(e2e): unbreak trtllm lanes broken by mcp 2.0.0a1 - #1675

fix(e2e): unbreak trtllm lanes broken by mcp 2.0.0a1#1675
slin1237 merged 1 commit into
mainfrom
fix/e2e-mcp-constraint

slin1237 commented Jun 11, 2026 •

edited by coderabbitai Bot

Loading

Uh oh!

coderabbitai Bot commented Jun 11, 2026 •

edited

Loading

Walkthrough

Changes

Estimated code review effort

Suggested labels

Suggested reviewers

Poem

❌ Failed checks (1 warning)

Uh oh!

gemini-code-assist Bot left a comment

Uh oh!

gemini-code-assist Bot Jun 11, 2026

Uh oh!

claude Bot commented Jun 11, 2026

Uh oh!

claude Bot left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

slin1237 commented Jun 11, 2026 • edited by coderabbitai Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Root cause (with receipts from job 80871944044)

Fix

Verification

Summary by CodeRabbit

Uh oh!

coderabbitai Bot commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Walkthrough

Changes

Estimated code review effort

Suggested labels

Suggested reviewers

Poem

❌ Failed checks (1 warning)

Uh oh!

gemini-code-assist Bot left a comment

Choose a reason for hiding this comment

Code Review

Uh oh!

gemini-code-assist Bot Jun 11, 2026

Choose a reason for hiding this comment

Uh oh!

claude Bot commented Jun 11, 2026

Uh oh!

claude Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

slin1237 commented Jun 11, 2026 •

edited by coderabbitai Bot

Loading

coderabbitai Bot commented Jun 11, 2026 •

edited

Loading