UPSTREAM PR #17460: Typo in json.gbnf

[loci-dev]

Mirrored from ggml-org/llama.cpp#17460

Small fix for typo in the number rule:

[eE] [-+]? [0-9] [1-9]{0,15}

This would allow one leading zero and no trailing zeros; e.g. forbidding e+10, e+100 etc. I assume the reverse was intended

[loci-review]

Explore the complete analysis inside the Version Insights

Pull Request #302 - Performance Review Summary

PR Title: UPSTREAM PR #17460: Typo in json.gbnf
Change Scope: Grammar definition file correction

---

Assessment

No performance impact detected. This PR corrects a typo in the JSON grammar definition file (grammars/json.gbnf) that fixes scientific notation exponent parsing. The change is a single-character substitution in a declarative grammar rule with zero measurable performance impact across all binaries.

Performance Metrics:

• Power consumption change: 0.0% across all 16 binaries
• Response Time change: No function-level deltas detected
• Throughput Time change: No function-level deltas detected
• Tokens per second impact: None (no inference path modifications)

Code Change Analysis:

The fix corrects the exponent pattern from [0-9] [1-9]{0,15} to [1-9] [0-9]{0,15}, which:

• Resolves incorrect rejection of valid exponents like e+10, e+100
• Properly rejects malformed exponents with leading zeros like e+01
• Aligns with JSON standard (RFC 8259)

Impact Scope:

• Affects only grammar-based constrained generation (structured output validation)
• Does not modify core inference functions: llama_decode, llama_encode, llama_tokenize
• No changes to Model Processing, Token Processing, Memory Management, or Batch Processing modules
• Grammar parsing occurs post-inference, outside the critical token generation path

Technical Correctness:

• Bug fix improves JSON validation accuracy
• No breaking changes for compliant JSON inputs
• Pattern matching complexity unchanged (identical character class operations)

---

Recommendation

Approve and merge. This is a correctness improvement with no performance implications. The change fixes a documented grammar bug while maintaining performance parity across all metrics. No action items required.

[loci-review]

Explore the complete analysis inside the Version Insights

Performance Analysis Summary - PR #302

Overview

PR #302 introduces a single-character grammar specification fix in grammars/json.gbnf to correct scientific notation parsing for JSON numbers. The change modifies the exponent pattern from [0-9] [1-9]{0,15} to [1-9] [0-9]{0,15}, enabling proper validation of multi-digit exponents containing zeros (e.g., e+10, e+100).

Performance analysis across all 16 binaries shows zero measurable impact. Power consumption remains stable at 226822 nJ for libllama.so, 285287 nJ for llama-tts, and 240164 nJ for llama-run. No functions within the Performance-Critical Areas (model processing, token processing, memory management, batch processing) were modified. This is a correctness fix that aligns the grammar with JSON RFC 8259 without affecting runtime execution paths or inference performance.