Bump version to 0.3.38

Signed-off-by: JamePeng <jame_peng@sina.com>

Author: JamePeng

CHANGELOG.md

@@ -7,6 +7,30 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.3.38] Optimized CJK Detokenization, Sync Grammar Parser, and Patched CUDA Graph Logs
+
+- perf: Optimize detokenize buffer sizing for CJK-heavy outputs
+    - Increase the initial detokenization buffer estimate from 1x tokens to `5x + 32` bytes. Performance analysis revealed that CJK-heavy outputs often require around 4.0x–5.04x bytes per token, with small-token edge cases reaching about 6.0x, so the previous estimate frequently forced a failed `llama_detokenize` call followed by a resize and retry. Previously, this resulted in almost twice as many calls to llama_detokenize.
+    - This reduces avoidable detokenization retries and cuts call overhead in CJK-heavy cases, resulting in an observed ~3–5% inference performance improvement.
+
+- patch(logger): filter out verbose noisy CUDA Graph debug logs
+    - Add a temporary patch in `ggml_log_callback` to suppress the noisy `CUDA Graph id %zu reused` messages generated by the underlying C++ backend. A complete logger refactoring is planned for better log control in the future.
+
+- docs: add LlamaGrammar wiki page and Update /docs/wiki/index.md
+    - Now Github Wiki Online: https://github.com/JamePeng/llama-cpp-python/wiki
+
+- feat(grammar): sync JSON schema to GBNF converter with upstream
+    - Allow `LlamaGrammar.from_json_schema` and `json_schema_to_gbnf` to accept both string and dict schema inputs.
+    - Expose `allow_fetch`, `dotall`, and `raw_pattern` arguments to the public API to match the upstream script.
+    - Fix missing handling for empty/unconstrained schema objects (e.g. `{"description": "..."}`) which now correctly default to accepting any value.
+    - Fix bug where `has_min and has_max` evaluated incorrectly when variables were zero. Replaced `!= None` with `is not None` in `_generate_min_max_int`.
+    - Update internal constants and regex patterns (`INVALID_RULE_CHARS_RE`, `GRAMMAR_LITERAL_ESCAPE_RE`, `GRAMMAR_RANGE_LITERAL_ESCAPE_RE`) to resolve character escaping issues.
+    - Update reference link to point to the new `ggml-org` organization.
+
+- feat: Update llama.cpp to [ggml-org/llama.cpp/commit/e48034dfc9e5705248fd39dc437ca887dc55a528](https://github.com/ggml-org/llama.cpp/commit/e48034dfc9e5705248fd39dc437ca887dc55a528)
+
+More information see: https://github.com/JamePeng/llama-cpp-python/compare/fe38cbfac424973ccd9cfaa199a64a02502af4d1...fe8657adf71856967320bfb932461d65ecf77b19
+
 ## [0.3.37] MoE CPU Offloading, O(1) Speculative Decoding, Thread-Safe Abort & New LLM Wiki
 
 - docs: A basic new documentation system for the LLM-Wiki has been initially established.

llama_cpp/__init__.py

@@ -1,4 +1,4 @@
 from .llama_cpp import *
 from .llama import *
 
-__version__ = "0.3.37"
+__version__ = "0.3.38"