chat : add Granite 4.0 chat template with correct tool_call role mapping#20804
Merged
pwilkin merged 4 commits intoApr 2, 2026
Merged
Conversation
Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
5 tasks
Collaborator
In general, this project doesn't accept fully generated PRs. For context, @jesus-talavera-ibm and I iterated on this on jesus-talavera-ibm#1, so there has been significant manual review of the generated code. |
This comment was marked as resolved.
This comment was marked as resolved.
gabe-l-hart
suggested changes
Mar 20, 2026
gabe-l-hart
left a comment
Collaborator
There was a problem hiding this comment.
I think we should refactor the test-chat-template.cpp changes to match the other tests, but otherwise this looks good!
| { "exaone-moe", LLM_CHAT_TEMPLATE_EXAONE_MOE }, | ||
| { "rwkv-world", LLM_CHAT_TEMPLATE_RWKV_WORLD }, | ||
| { "granite", LLM_CHAT_TEMPLATE_GRANITE }, | ||
| { "granite", LLM_CHAT_TEMPLATE_GRANITE_3_X }, |
Collaborator
There was a problem hiding this comment.
The goal of this change is to carefully delineate between template for the 3.x series and the 4.0 series (there may be further 4.x template changes in the future, thus being explicit about 4_0).
As written, this is nicely backwards-compatible since the enum name is only used internally and the string name is mapped consistently.
| } | ||
| } | ||
|
|
||
| // Test Granite 3.x template (LLM_CHAT_TEMPLATE_GRANITE_3_X) — backwards compatibility |
Collaborator
There was a problem hiding this comment.
I should have caught this in pre-review, but I think we should try to rewrite these tests to match the ones above that all follow a standard format and use common test logic.
Contributor
Author
There was a problem hiding this comment.
Thank you, Gabe! I've adapted the tests here: 3f2a48a
gabe-l-hart
approved these changes
Apr 1, 2026
gabe-l-hart
left a comment
Collaborator
There was a problem hiding this comment.
I think this change looks good.
@pwilkin @ngxson @ggerganov We've validated this against both 3.x and 4.0 models and the code/tests look tight, so I think it's ready for CI and maintainer review.
pwilkin
approved these changes
Apr 1, 2026
ggerganov
approved these changes
Apr 2, 2026
slartibardfast
pushed a commit
to slartibardfast/llama.cpp
that referenced
this pull request
Apr 12, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Seunghhon
pushed a commit
to Seunghhon/llama.cpp
that referenced
this pull request
Apr 26, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
rsenthilkumar6
pushed a commit
to rsenthilkumar6/llama.cpp
that referenced
this pull request
May 1, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
ljubomirj
pushed a commit
to ljubomirj/llama.cpp
that referenced
this pull request
May 6, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
my-other-github-account
pushed a commit
to my-other-github-account/llama.cpp
that referenced
this pull request
May 15, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
my-other-github-account
pushed a commit
to my-other-github-account/llama.cpp
that referenced
this pull request
May 15, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
fewtarius
pushed a commit
to fewtarius/llama.cpp
that referenced
this pull request
May 30, 2026
…ing (ggml-org#20804) * chat : add Granite 4.0 chat template with correct tool_call role mapping Introduce `LLM_CHAT_TEMPLATE_GRANITE_4_0` alongside the existing Granite 3.x template (renamed `LLM_CHAT_TEMPLATE_GRANITE_3_X`). The Granite 4.0 Jinja template uses `<tool_call>` XML tags and maps the `assistant_tool_call` role to `<|start_of_role|>assistant<|end_of_role|><|tool_call|>`. Without a matching C++ handler, the fallback path emits the literal role `assistant_tool_call` which the model does not recognize, breaking tool calling when `--jinja` is not used. Changes: - Rename `LLM_CHAT_TEMPLATE_GRANITE` to `LLM_CHAT_TEMPLATE_GRANITE_3_X` (preserves existing 3.x behavior unchanged) - Add `LLM_CHAT_TEMPLATE_GRANITE_4_0` enum, map entry, and handler - Detection: `<|start_of_role|>` + (`<tool_call>` or `<tools>`) → 4.0, otherwise → 3.x - Add production Granite 4.0 Jinja template - Add tests for both 3.x and 4.0 template paths (C++ and Jinja) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * Code review: follow standard format and use common logic in test-chat-template.cpp * Rename custom_conversation variable for extra_conversation to give it a more meaningful name --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
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
4 participants
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.
Summary
Add
LLM_CHAT_TEMPLATE_GRANITE_4_0as a parallel path to the existing Granite 3.x template, fixingassistant_tool_callrole mapping for Granite 4.0 models when--jinjais not used.Problem
The Granite 4.0 Jinja template maps
assistant_tool_callto<|start_of_role|>assistant<|end_of_role|><|tool_call|>, but the C++ template handler emits the literal role<|start_of_role|>assistant_tool_call<|end_of_role|>which the model does not recognize. This silently breaks tool calling on the C++ template path (without--jinja).Changes
LLM_CHAT_TEMPLATE_GRANITE→LLM_CHAT_TEMPLATE_GRANITE_3_X(preserves existing 3.x behavior unchanged)LLM_CHAT_TEMPLATE_GRANITE_4_0enum, map entry ("granite-4.0"), and handler<|start_of_role|>+ (<tool_call>or<tools>) → 4.0; otherwise → 3.xtest-chat-template.cpptest-chat.cppFiles changed
src/llama-chat.hGRANITE_3_X/GRANITE_4_0enum valuessrc/llama-chat.cppmodels/templates/ibm-granite-granite-4.0.jinjatests/test-chat-template.cpptests/test-chat.cppTest plan
test-chat-template— Granite 3.x backwards compat (C++ + Jinja): PASStest-chat-template— Granite 4.0 fix (C++ + Jinja): PASStest-chat— Granite 3.3 peg parser (reasoning + basic chat): PASStest-chat— Granite 4.0 peg parser (basic chat + tool calling): PASS🤖 Generated with Claude Code
EDIT: I used Claude Code as start point and then iterated from it manually. You can check the whole process here: jesus-talavera-ibm#1