UPSTREAM PR #17330: Fix grammar README: Update non-terminal naming rules

[DajanaV]

Mirrored from ggml-org/llama.cpp#17330

I noticed the grammar README documentation was outdated. It stated that rule names must be 'dashed lowercase words', but when I looked at the examples like c.gbnf, I saw it uses dataType with an uppercase letter.

After checking the discussion in #7720, it's clear the parser actually supports much more flexibility than the docs suggested:

• Uppercase and lowercase letters
• Numbers
• Both dashes and underscores

This PR updates the documentation to accurately reflect what the parser actually supports. The examples work fine - the docs just needed updating to match reality.

Changes:

• Updated the 'Non-Terminals and Terminals' section to list all supported characters
• Added examples showing the flexibility (dataType, UPPER-CASE, rule_123)

This should help people understand they have more flexibility when naming their grammar rules.

Fixes #7720

[loci-agentic-ai]

Access the complete analysis in the LOCI Dashboard

Performance Analysis Summary

Based on the comprehensive analysis of PR #243 in the llama.cpp repository, this change represents a documentation-only modification with no impact on executable code or performance metrics.

Change Overview

The pull request updates the grammar README file (grammars/README.md) to correct outdated documentation about non-terminal naming rules. The change clarifies that grammar rule names support uppercase letters, lowercase letters, numbers, dashes, and underscores, expanding beyond the previously documented "dashed lowercase words" restriction.

Scope of Changes:

• Single file modification: grammars/README.md
• Net change: 0 lines (1 addition, 1 deletion)
• Change type: Text clarification and example expansion

Performance Impact Assessment

Function-Level Analysis:

• No functions modified or affected
• No changes to Response Time or Throughput metrics
• No impact on critical inference functions (llama_decode, llama_encode, llama_tokenize)
• No changes to core modules (Model Processing, Token Processing, Memory Management, Batch Processing)

System Performance:

• Tokens per second: No impact (no executable code changes)
• Power consumption: No changes across any binaries
• Memory usage: No modifications to memory management or KV cache systems
• Inference pipeline: No alterations to tokenization or model processing paths

Technical Validation

The documentation update correctly aligns with the actual grammar parser implementation, which already supported the expanded naming conventions. The change improves developer experience by providing accurate guidance and concrete examples without affecting runtime behavior.

Risk Assessment:

• No regression potential (documentation only)
• No breaking changes to existing functionality
• No impact on build process or dependencies

This change represents a pure documentation improvement that enhances accuracy and usability without any performance implications for the llama.cpp inference engine.