Bump DiffusionGemma #24423 pin to c3fb9724 (Add tool calling)#32
Conversation
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: a6bff50ef2
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| ], | ||
| "prs": [ | ||
| "https://github.com/ggml-org/llama.cpp/pull/24423/commits/1d2faac2d8248c3d5df999712c42fe0fc72e7f07", | ||
| "https://github.com/ggml-org/llama.cpp/pull/24423/commits/c3fb97241295c196e09b783e705e84b96cd1bd74", |
There was a problem hiding this comment.
Pin a commit that ignores tool_choice
This updated pin pulls in c3fb972, whose only tool-calling change reads req["tools"] into the chat-template inputs but never reads tool_choice. For OpenAI-compatible requests that include tools while setting tool_choice: "none", the diffusion visual server still exposes those tools to the model, so callers cannot reliably suppress tool calls in the same way the regular server does; please pin a follow-up that handles tool_choice or avoid advertising this as the tool-calling fix until that path is covered.
Useful? React with 👍 / 👎.
| ], | ||
| "prs": [ | ||
| "https://github.com/ggml-org/llama.cpp/pull/24423/commits/1d2faac2d8248c3d5df999712c42fe0fc72e7f07", | ||
| "https://github.com/ggml-org/llama.cpp/pull/24423/commits/c3fb97241295c196e09b783e705e84b96cd1bd74", |
There was a problem hiding this comment.
Preserve tool-call grammar from the pinned commit
The pinned c3fb972 path only uses common_chat_templates_apply(...).prompt after adding inputs.tools; it discards the grammar and grammar_lazy that common_chat_templates_apply returns and that tools/server/server-common.cpp forwards for tool requests. For tools with required fields, enums, or numeric schemas, the visual server can now render the tools but still samples unconstrained text, so callers can receive malformed or out-of-schema tool calls instead of the schema-correct calls expected from OpenAI-compatible tool use.
Useful? React with 👍 / 👎.
Bump the DiffusionGemma pin (ggml-org#24423) from
1d2faactoc3fb9724.Why
The current pin (
1d2faac, "Fix merge conflicts", 2026-07-03) is one commit behindc3fb9724("Add tool calling", 2026-07-04), which is the head of ggml-org#24423. That commit adds the three lines that let the visual server read tools from the request:Without it the shipped
llama-diffusion-gemma-visual-servernever renders the caller's tools into the chat template, so DiffusionGemma cannot produce schema-correct tool calls. This is the binary half of the DiffusionGemma tool-calling path that unslothai/unsloth#6851 and unslothai/unsloth-zoo#864 depend on.Verification
Built
c3fb9724from source (single-arch CUDA sm_100) and ran it under Studio withunsloth/diffusiongemma-26B-A4B-it-GGUF(Q4_K_M):1d2faac): a plain OpenAI tools request returned "I do not have access to ... tools" (the model never saw the tool).c3fb9724): the same request returns a structuredtool_callsarray,get_weather(location="San Francisco").tool_choice: "none"correctly suppresses the tool.Plain chat and the diffusion visualization are unchanged.
Lint: the new URL matches the pr-set schema and
c3fb9724is the current head of ggml-org#24423.