vllm addition to torchtalk#10
Open
arkadip-maitra wants to merge 5 commits into
Open
Conversation
arkadip-maitra
changed the title
Restore the hybrid-discovery vLLM graph fixes so TorchTalk stays resilient to current vLLM layout drift, and add a clean LLM + TorchTalk(vLLM) vs LLM-only benchmark package that shows the query quality, outputs, and summary results.
Apply Ruff formatting to the vLLM index builder so the src/torchtalk format check passes in CI.
llm models are gpt 5.4 and claude opus 4.6
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #5
Extend TorchTalk with first-class vllm support through a framework adapter architecture, while preserving PyTorch as the default path. This adds a real VllmAdapter, static vLLM indexing across API/engine/registry/IR/native-binding layers, condition-aware graph/proof traces, and framework-aware CLI/MCP behavior.
Tested thoroughly across both repos: TorchTalk lint passed, the full TorchTalk suite passed with
PYTORCH_SOURCE=/data/pytorch VLLM_SOURCE=/data/vllm(487 tests), and the new vLLM-focused TorchTalk integration/quality suites all passed (45 tests) covering adapter bootstrap, CLI paths, MCP-facingget_status / search / trace / graph, unsupported-tool behavior, entity-count thresholds, proof traces, and conditional graph evidence. I also source-built vLLM in a fresh conda env, verified imports/native bindings, ran focused vLLM native tests (1508 passed), and confirmed real runtime behavior on H200 with offline generation, pooling/embeddings, isolated OpenAI-compatible serve, and a 2-GPU tensor-parallel serve smoke.