PyPI - llguidance - Versions diffs - 1.1.1__tar.gz → 1.2.0__tar.gz - Mend

llguidance 1.1.1tar.gz → 1.2.0tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (193) hide show

llguidance-1.2.0/.github/workflows/code-coverage.yml ADDED Viewed

@@ -0,0 +1,62 @@
+name: Code Coverage
+permissions:
+  contents: read
+on:
+  pull_request:
+  push:
+    branches: [ "main" ]
+  workflow_dispatch:
+    inputs:
+      commit_id:
+        description: 'Branch or Commit ID (optional)'
+        required: false
+        type: string
+env:
+  CARGO_TERM_COLOR: always
+  RUSTFLAGS: "-Cinstrument-coverage"
+  LLVM_PROFILE_FILE: "llg-%p-%m.profraw"
+jobs:
+  code-cov:
+    runs-on: ubuntu-latest
+    steps:
+    - name: Checkout repo at ${{ github.event_name == 'workflow_dispatch' && inputs.commit_id || github.sha }}
+      uses: actions/checkout@v4
+      with:
+        ref: ${{ github.event_name == 'workflow_dispatch' && inputs.commit_id || github.sha }}
+    - name: Update toolchain
+      run: rustup component add llvm-tools
+    - name: Install grcov
+      run: cargo install grcov
+    - name: Build everything
+      run: cargo build
+    - name: Run tests
+      run: cargo test
+    - name: Check environment
+      run: |
+        echo "CARGO_TERM_COLOR: $CARGO_TERM_COLOR"
+        echo "RUSTFLAGS: $RUSTFLAGS"
+        echo "LLVM_PROFILE_FILE: $LLVM_PROFILE_FILE"
+    - name: Generate coverage report
+      run: |
+        grcov . -s . --binary-path target/debug/ -t html --branch --ignore-not-existing -o target/debug/coverage/
+    - name: Check output
+      run: ls target/debug/coverage/
+    - uses: actions/upload-artifact@v4
+      with:
+        name: coverage-report
+        path: target/debug/coverage/

{llguidance-1.1.1 → llguidance-1.2.0}/CHANGELOG.md RENAMED Viewed

@@ -4,109 +4,121 @@ All notable changes to this project will be documented in this file. Dates are d
 If a release doesn't introduce any interesting changes (build fixes etc.), it's skipped.
-#### [1.1.1](https://github.com/guidance-ai/llguidance/compare/v1.1.0...1.1.1) 2025-07-23
+#### [v1.1.3](https://github.com/guidance-ai/llguidance/compare/v1.1.2...v1.1.3) 2025-08-12
+- support multithreaded compute bitmask for speculative decoding [`#225`](https://github.com/guidance-ai/llguidance/pull/225)
+  - thank you [@ZonePG](https://github.com/ZonePG)!
+- `force_lexeme_end` -> `try_lexeme_end` in lark lexer when out of input [`#229`](https://github.com/guidance-ai/llguidance/pull/229); fixes [`#228`](https://github.com/guidance-ai/llguidance/issues/228)
+- more JSON test coverage
+#### [v1.1.2](https://github.com/guidance-ai/llguidance/compare/v1.1.1...v1.1.2) 2025-08-08
+- add flag in ParserLimits to disable verbose errors [`#227`](https://github.com/guidance-ai/llguidance/pull/227)
+- new tests and cleanups
+#### [v1.1.1](https://github.com/guidance-ai/llguidance/compare/v1.1.0...v1.1.1) 2025-07-23
 - prevent error state when calling `try_consume_tokens` after parser is stopped [`#213`](https://github.com/guidance-ai/llguidance/pull/213); fixes [`#211`](https://github.com/guidance-ai/llguidance/issues/211)
 - set parser stop condition in `try_consume_tokens` even when some tokens are rejected [`#212`](https://github.com/guidance-ai/llguidance/pull/212)
-#### [1.1.0](https://github.com/guidance-ai/llguidance/compare/v1.0.1...1.1.0) 2025-07-18
+#### [v1.1.0](https://github.com/guidance-ai/llguidance/compare/v1.0.1...v1.1.0) 2025-07-18
 - disable hf tokenizer truncation and padding [`#205`](https://github.com/guidance-ai/llguidance/pull/205); fixes [`#1322`](https://github.com/guidance-ai/guidance/issues/1322)
 - llama_cpp tokenizers: infer added tokens starting/ending with &lt; and &gt; to be special tokens [`#202`](https://github.com/guidance-ai/llguidance/pull/202)
 - add lark syntax for "any token" and negation of token ranges [`#201`](https://github.com/guidance-ai/llguidance/pull/201)
 - add de-recursion cook book to docs [`#199`](https://github.com/guidance-ai/llguidance/pull/199)
-#### [1.0.1](https://github.com/guidance-ai/llguidance/compare/v1.0.0...1.0.1) 2025-07-03
+#### [v1.0.1](https://github.com/guidance-ai/llguidance/compare/v1.0.0...v1.0.1) 2025-07-03
 - fix: tokenizers normalizers sequence api changed [`#195`](https://github.com/guidance-ai/llguidance/pull/195)
 - Strip debug info from the wheels [`#194`](https://github.com/guidance-ai/llguidance/pull/194)
 Thank you @ammar-elsabe and @Ahajha!
-#### [1.0.0](https://github.com/guidance-ai/llguidance/compare/v0.7.30...1.0.0) 2025-06-23
+#### [v1.0.0](https://github.com/guidance-ai/llguidance/compare/v0.7.30...v1.0.0) 2025-06-23
 This is identical to `0.7.30`, but indicates intended stability and from now on we'll try to follow semver.
-#### [0.7.30](https://github.com/guidance-ai/llguidance/compare/v0.7.29...0.7.30) 2025-06-23
+#### [v0.7.30](https://github.com/guidance-ai/llguidance/compare/v0.7.29...v0.7.30) 2025-06-23
 - parametric grammars [`#192`](https://github.com/guidance-ai/llguidance/pull/192)
 - allow for tokens up to ~2k bytes; fixes #188 [`#188`](https://github.com/guidance-ai/llguidance/issues/188)
-#### [0.7.29](https://github.com/guidance-ai/llguidance/compare/v0.7.28...0.7.29) 2025-06-06
+#### [v0.7.29](https://github.com/guidance-ai/llguidance/compare/v0.7.28...v0.7.29) 2025-06-06
 - cargo fmt
-#### [0.7.28](https://github.com/guidance-ai/llguidance/compare/v0.7.27...0.7.28) 2025-06-06
+#### [v0.7.28](https://github.com/guidance-ai/llguidance/compare/v0.7.27...v0.7.28) 2025-06-06
 - fix lexer_stack=... panic with numeric tokens [`4e91b0f`](https://github.com/guidance-ai/llguidance/commit/4e91b0fa0c03572a5fc221ac0e0b05035af9dcfa)
-#### [0.7.27](https://github.com/guidance-ai/llguidance/compare/v0.7.26...0.7.27) 2025-06-04
+#### [v0.7.27](https://github.com/guidance-ai/llguidance/compare/v0.7.26...v0.7.27) 2025-06-04
 - add toktrie_tiktoken and llguidance.tiktoken.lltokenizer_from_encoding [`#154`](https://github.com/guidance-ai/llguidance/issues/154)
 - implement clone on StopController [`#185`](https://github.com/guidance-ai/llguidance/issues/185)
-#### [0.7.26](https://github.com/guidance-ai/llguidance/compare/v0.7.25...0.7.26) 2025-05-30
+#### [v0.7.26](https://github.com/guidance-ai/llguidance/compare/v0.7.25...v0.7.26) 2025-05-30
 - add support for & and ~ in lark regexes [`96fcee3`](https://github.com/guidance-ai/llguidance/commit/96fcee373697b57bead94d1bc06c17cf1c6134e4)
 - dump grammar in errors in LLInterpreter [`#183`](https://github.com/guidance-ai/llguidance/pull/183)
 - don't check lexer bytes invariant when we cannot rollback [`ec22083`](https://github.com/guidance-ai/llguidance/commit/ec220837051513a70177974ca389b7bf387455f1)
-#### [0.7.25](https://github.com/guidance-ai/llguidance/compare/v0.7.24...0.7.25) 2025-05-28
+#### [v0.7.25](https://github.com/guidance-ai/llguidance/compare/v0.7.24...v0.7.25) 2025-05-28
 - add parse_special=False to tokenize_str/bytes() in python [`#181`](https://github.com/guidance-ai/llguidance/pull/181)
-#### [0.7.24](https://github.com/guidance-ai/llguidance/compare/v0.7.23...0.7.24) 2025-05-23
+#### [v0.7.24](https://github.com/guidance-ai/llguidance/compare/v0.7.23...v0.7.24) 2025-05-23
 - add the sentinel token hack, fixes #180 [`#180`](https://github.com/guidance-ai/llguidance/issues/180)
-#### [0.7.23](https://github.com/guidance-ai/llguidance/compare/v0.7.22...0.7.23) 2025-05-22
+#### [v0.7.23](https://github.com/guidance-ai/llguidance/compare/v0.7.22...v0.7.23) 2025-05-22
 - native llama.cpp tokenizer support [`#179`](https://github.com/guidance-ai/llguidance/pull/179)
 - improve special token detection in HF tokenizers [`6cae393`](https://github.com/guidance-ai/llguidance/commit/6cae393b9c04fe67621615ff22b46beab512d069)
-#### [0.7.22](https://github.com/guidance-ai/llguidance/compare/v0.7.21...0.7.22) 2025-05-21
+#### [v0.7.22](https://github.com/guidance-ai/llguidance/compare/v0.7.21...v0.7.22) 2025-05-21
 - Keep EOS token bytes in `TokenizerWrapper` [`#178`](https://github.com/guidance-ai/llguidance/pull/178)
 - Stop using prefix/sentinel strings for `TokenizerWrapper` [`#175`](https://github.com/guidance-ai/llguidance/pull/175)
 - avoid taking poisoned locks, see [`#174`](https://github.com/guidance-ai/llguidance/issues/174) [`d41aa9a`](https://github.com/guidance-ai/llguidance/commit/d41aa9a4427967708a951506b2bc0e395871b6c8); thanks [@g-eoj](https://github.com/g-eoj)
-#### [0.7.21](https://github.com/guidance-ai/llguidance/compare/v0.7.20...0.7.21) 2025-05-20
+#### [v0.7.21](https://github.com/guidance-ai/llguidance/compare/v0.7.20...v0.7.21) 2025-05-20
 - include parser state in errors [`82e34da`](https://github.com/guidance-ai/llguidance/commit/82e34da704d22f04979d8cbc54a0ac00885a277d)
 - tighten email format in JSON schema [`7454ea9`](https://github.com/guidance-ai/llguidance/commit/7454ea9df958f8bcc42e6bb986d6de397de65b3e)
-#### [0.7.20](https://github.com/guidance-ai/llguidance/compare/v0.7.19...0.7.20) 2025-05-15
+#### [v0.7.20](https://github.com/guidance-ai/llguidance/compare/v0.7.19...v0.7.20) 2025-05-15
 - use fancy-regex instead of onig as tokenizers regex library [`#172`](https://github.com/guidance-ai/llguidance/pull/172)
   - fixes compilation on GCC 15, thanks [@Slowki](https://github.com/Slowki)
 - msrv 1.80 support (incl. derivre bump) [`c89e386`](https://github.com/guidance-ai/llguidance/commit/c89e386685cd911a89fd47df225de88f88c10883), thank you [@nteodosio](https://github.com/nteodosio) for initial [PR](https://github.com/guidance-ai/llguidance/pull/170)!
-#### [0.7.19](https://github.com/guidance-ai/llguidance/compare/v0.7.18...0.7.19) 2025-04-24
+#### [v0.7.19](https://github.com/guidance-ai/llguidance/compare/v0.7.18...v0.7.19) 2025-04-24
 - fix a numeric token bug [`1f59edf`](https://github.com/guidance-ai/llguidance/commit/1f59edfc49b44cfba74b2380f34874a0778d9441)
-#### [0.7.18](https://github.com/guidance-ai/llguidance/compare/v0.7.17...0.7.18) 2025-04-22
+#### [v0.7.18](https://github.com/guidance-ai/llguidance/compare/v0.7.17...v0.7.18) 2025-04-22
 - apply x-guidance also in %json{} [`2627891`](https://github.com/guidance-ai/llguidance/commit/2627891c72c7e38062cd3e052f1de146d2e21635)
 - more sensible llg_validate_grammar() signature [`41928c0`](https://github.com/guidance-ai/llguidance/commit/41928c07298e69e3c8adc4a3c1f43ef9b1cc1c6b)
-#### [0.7.17](https://github.com/guidance-ai/llguidance/compare/v0.7.16...0.7.17) 2025-04-22
+#### [v0.7.17](https://github.com/guidance-ai/llguidance/compare/v0.7.16...v0.7.17) 2025-04-22
 - support for min/maxProperties in JSON Schema [`#168`](https://github.com/guidance-ai/llguidance/issues/168)
 - give priority to &lt;[123]&gt; over "foo" in grammar [`3e9f3b5`](https://github.com/guidance-ai/llguidance/commit/3e9f3b5e8c1cac92daab6e9709f01ebccc20342b)
-#### [0.7.16](https://github.com/guidance-ai/llguidance/compare/v0.7.15...0.7.16) 2025-04-17
+#### [v0.7.16](https://github.com/guidance-ai/llguidance/compare/v0.7.15...v0.7.16) 2025-04-17
 - fix special token tokenization [`ae7870f`](https://github.com/guidance-ai/llguidance/commit/ae7870f05ca0de68599088607ba742b7071f92ad)
-#### [0.7.15](https://github.com/guidance-ai/llguidance/compare/v0.7.14...0.7.15) 2025-04-16
+#### [v0.7.15](https://github.com/guidance-ai/llguidance/compare/v0.7.14...v0.7.15) 2025-04-16
 - support for patternProperties in JSON schema [`#167`](https://github.com/guidance-ai/llguidance/pull/167)
 - add lenient option to JSON schemas [`#163`](https://github.com/guidance-ai/llguidance/pull/163) [`#136`](https://github.com/guidance-ai/llguidance/issues/136)
 - Add llg_validate_grammar() in C FFI [`e5c21cf`](https://github.com/guidance-ai/llguidance/commit/e5c21cf480a17e6b310e46b24b272576cfd9c4c6)
-#### [0.7.14](https://github.com/guidance-ai/llguidance/compare/v0.7.13...0.7.14) 2025-04-11
+#### [v0.7.14](https://github.com/guidance-ai/llguidance/compare/v0.7.13...v0.7.14) 2025-04-11
 - support %lark { ... } syntax for nested grammars [`#157`](https://github.com/guidance-ai/llguidance/pull/157)
 - treat \d and \w in json schema as ASCII; fix ^$ anchors [`#158`](https://github.com/guidance-ai/llguidance/issues/158)
@@ -115,19 +127,19 @@ This is identical to `0.7.30`, but indicates intended stability and from now on
 - expose regex_to_lark() in Rust and Python; add \d\w\s replacement [`78fb32f`](https://github.com/guidance-ai/llguidance/commit/78fb32fe2745d30ca94a62b00e5a7299750d80b0)
 - fix usage of / vs \* in python signatures [`ca73c2a`](https://github.com/guidance-ai/llguidance/commit/ca73c2abd44e75d569230b942f53c72b052ed2ab)
-#### [0.7.13](https://github.com/guidance-ai/llguidance/compare/v0.7.12...0.7.13) 2025-04-05
+#### [v0.7.13](https://github.com/guidance-ai/llguidance/compare/v0.7.12...v0.7.13) 2025-04-05
 - expose LLParserLimits in Python API [`598dc8f`](https://github.com/guidance-ai/llguidance/commit/598dc8f37f69f51244e54d9885445abf02a515a7)
 - pre-compute lexer states for particularly large regexes (can be disabled in ParserLimits)
-#### [0.7.12](https://github.com/guidance-ai/llguidance/compare/v0.7.11...0.7.12) 2025-04-04
+#### [v0.7.12](https://github.com/guidance-ai/llguidance/compare/v0.7.11...v0.7.12) 2025-04-04
 - performance optimizations
 - use factory in C FFI (otherwise slicer was not used)
 - add some null checks and safety comments in C FFI
 - implement subgrammar lexeme class merging; fixes [`#113`](https://github.com/guidance-ai/llguidance/issues/113)
-#### [0.7.11](https://github.com/guidance-ai/llguidance/compare/v0.7.10...0.7.11) 2025-03-27
+#### [v0.7.11](https://github.com/guidance-ai/llguidance/compare/v0.7.10...v0.7.11) 2025-03-27
 - add StructTag python API; fixes [`#146`](https://github.com/guidance-ai/llguidance/issues/146)
 - fix handling of AddedToken.special (gemma tokenizer, fixes [`#147`](https://github.com/guidance-ai/llguidance/issues/147))

{llguidance-1.1.1 → llguidance-1.2.0}/Cargo.lock RENAMED Viewed

@@ -1241,7 +1241,7 @@ checksum = "241eaef5fd12c88705a01fc1066c48c4b36e0dd4377dcdc7ec3942cea7a69956"
 [[package]]
 name = "llguidance"
-version = "1.1.1"
+version = "1.2.0"
 dependencies = [
  "anyhow",
  "derivre",
@@ -1260,7 +1260,7 @@ dependencies = [
 [[package]]
 name = "llguidance_py"
-version = "1.1.1"
+version = "1.2.0"
 dependencies = [
  "anyhow",
  "bytemuck",
@@ -2478,7 +2478,7 @@ dependencies = [
 [[package]]
 name = "toktrie"
-version = "1.1.1"
+version = "1.2.0"
 dependencies = [
  "anyhow",
  "bytemuck",
@@ -2489,7 +2489,7 @@ dependencies = [
 [[package]]
 name = "toktrie_hf_downloader"
-version = "1.1.1"
+version = "1.2.0"
 dependencies = [
  "anyhow",
  "hf-hub",
@@ -2500,7 +2500,7 @@ dependencies = [
 [[package]]
 name = "toktrie_hf_tokenizers"
-version = "1.1.1"
+version = "1.2.0"
 dependencies = [
  "anyhow",
  "log",
@@ -2512,7 +2512,7 @@ dependencies = [
 [[package]]
 name = "toktrie_tiktoken"
-version = "1.1.1"
+version = "1.2.0"
 dependencies = [
  "anyhow",
  "log",

{llguidance-1.1.1 → llguidance-1.2.0}/PKG-INFO RENAMED Viewed

@@ -1,10 +1,10 @@
 Metadata-Version: 2.4
 Name: llguidance
-Version: 1.1.1
+Version: 1.2.0
 License-File: LICENSE
 Summary: Bindings for the Low-level Guidance (llguidance) Rust library for use within Guidance
 Author: Michal Moskal
-License: MIT
+License-Expression: MIT
 Requires-Python: >=3.9
 Description-Content-Type: text/markdown; charset=UTF-8; variant=GFM
 Project-URL: repository, https://github.com/microsoft/llguidance

{llguidance-1.1.1 → llguidance-1.2.0}/parser/Cargo.toml RENAMED Viewed

@@ -1,6 +1,6 @@
 [package]
 name = "llguidance"
-version = "1.1.1"
+version = "1.2.0"
 edition = "2021"
 license = "MIT"
 description = "Super-fast Structured Outputs"

{llguidance-1.1.1 → llguidance-1.2.0}/parser/llguidance.h RENAMED Viewed

@@ -78,6 +78,12 @@ typedef struct LlgParserLimits {
    * Default: true
    */
   bool precompute_large_lexemes;
+  /**
+   * If true, include parser state (including tokens so far) and grammar in
+   * errors.
+   * Default: true
+   */
+  bool verbose_errors;
 } LlgParserLimits;
 typedef struct LlgConstraintInit {

{llguidance-1.1.1 → llguidance-1.2.0}/parser/src/api.rs RENAMED Viewed

@@ -258,6 +258,11 @@ pub struct ParserLimits {
     /// the time it takes to construct the lexer.
     /// Default: true
     pub precompute_large_lexemes: bool,
+    /// If true, include parser state (including tokens so far) and grammar in
+    /// errors.
+    /// Default: true
+    pub verbose_errors: bool,
 }
 impl Default for ParserLimits {
@@ -270,6 +275,7 @@ impl Default for ParserLimits {
             max_grammar_size: 500_000,     // fhir schema => 200k
             step_max_items: 50_000,        //
             precompute_large_lexemes: true,
+            verbose_errors: true,
         }
     }
 }

{llguidance-1.1.1 → llguidance-1.2.0}/parser/src/lark/lexer.rs RENAMED Viewed

@@ -279,7 +279,7 @@ pub fn lex_lark(input: &str) -> Result<Vec<Lexeme>> {
     while idx <= input_bytes.len() {
         let mut b = b'\n';
         let res = if idx == input_bytes.len() {
-            lexer.force_lexeme_end(state)
+            lexer.try_lexeme_end(state)
         } else {
             b = input_bytes[idx];
             lexer.advance(state, b, false)

{llguidance-1.1.1 → llguidance-1.2.0}/parser/src/panic_utils.rs RENAMED Viewed

@@ -24,7 +24,7 @@ pub fn mk_panic_error(info: &Box<dyn Any + Send>) -> String {
     let b = BACKTRACE.with(|b| b.take());
     if let Some(b) = b {
-        format!("panic: {msg}\n{b}")
+        format!("panic: {msg}\n<backtrace>\n{b}\n</backtrace>")
     } else {
         format!("panic: {msg}")
     }

{llguidance-1.1.1 → llguidance-1.2.0}/parser/src/tokenparser.rs RENAMED Viewed

@@ -280,11 +280,15 @@ impl TokenParser {
     }
     pub fn augment_err(&self, e: impl Display) -> String {
-        format!(
-            "{e}\n<state>\n{}\n</state><grammar>\n{}\n</grammar>",
-            self.dump_state(),
-            self.dbg_grammar
-        )
+        if self.limits.verbose_errors {
+            format!(
+                "{e}\n<state>\n{}\n</state><grammar>\n{}\n</grammar>",
+                self.dump_state(),
+                self.dbg_grammar
+            )
+        } else {
+            format!("{e}\n<non-verbose/>")
+        }
     }
     pub fn dump_state(&self) -> String {

{llguidance-1.1.1 → llguidance-1.2.0}/pyproject.toml RENAMED Viewed

@@ -1,6 +1,6 @@
 [project]
 name = "llguidance"
-version = "1.1.1"
+version = "1.2.0"
 description = "Bindings for the Low-level Guidance (llguidance) Rust library for use within Guidance"
 requires-python = ">=3.9"
 license = "MIT"

{llguidance-1.1.1 → llguidance-1.2.0}/python/llguidance/_lib.pyi RENAMED Viewed

@@ -526,18 +526,49 @@ class LLExecutor:
         self,
         interpreters: List[Tuple[LLMatcher, int]],
         trg_pointer: int,
-        one_mask_byte_size: int,
+        one_mask_bytes: int,
         trg_batch_size: int,
     ) -> None:
         """
         Compute the token mask directly into memory at the specified pointer.
         For each matcher, provide the index of the target mask.
-        If index is K, the memory will be written at trg_pointer + K * one_mask_byte_size,
+        If index is K, the memory will be written at trg_pointer + K * one_mask_bytes,
         where K < trg_batch_size.
-        Memory has to have size trg_batch_size * one_mask_byte_size.
+        Memory has to have size trg_batch_size * one_mask_bytes.
         Prefer to use fill_next_token_bitmask_par(), which wraps this.
         """
+    def unsafe_compute_mask_ptr_with_draft_token(
+        self,
+        interpreters: List[Tuple[LLMatcher, int, List[int]]],
+        trg_pointer: int,
+        one_mask_bytes: int,
+        trg_batch_size: int,
+    ) -> None:
+        """
+        Compute the token mask directly into memory at the specified pointer, including draft tokens.
+        This function extends unsafe_compute_mask_ptr() to handle draft tokens in speculative decoding.
+        For each matcher in the batch, it computes masks for both the current position and all draft tokens.
+        Args:
+            interpreters: List of tuples containing:
+                - LLMatcher: The matcher object for constrained generation
+                - int: Index K indicating the target mask position (K < trg_batch_size)
+                - List[int]: Draft tokens to be processed for speculative decoding
+            trg_pointer: Memory address where mask data will be written
+            one_mask_bytes: Size in bytes of a single token mask
+            trg_batch_size: Total batch size for memory allocation validation
+        Memory Layout:
+            - Main mask written at: trg_pointer + K * one_mask_bytes
+            - Draft token i mask written at: trg_pointer + (K + i + 1) * one_mask_bytes
+            - Total memory required: trg_batch_size * one_mask_bytes
+        The function processes each matcher's draft tokens sequentially, advancing the matcher state
+        for each valid token until encountering an invalid token or termination condition.
+        State rollback is performed to maintain matcher consistency.
+        """
 class JsonCompileOptions(TypedDict, total=False):
     # defaults to ","
@@ -565,6 +596,7 @@ class LLParserLimits:
         max_lexer_states: Optional[int] = None,
         max_grammar_size: Optional[int] = None,
         precompute_large_lexemes: Optional[bool] = None,
+        verbose_errors: Optional[bool] = None,
     ) -> None:
         """
         ParserLimits configuration for controlling parser and lexer resource usage.
@@ -597,6 +629,10 @@ class LLParserLimits:
             precompute_large_lexemes (Optional[bool]):
                 Whether to run large regexes eagerly on the entire token trie during lexer build.
                 Increases lexer construction time, but speeds up mask computation. Default: True.
+            verbose_errors (Optional[bool]):
+                If true, include parser state and grammar details in error messages.
+                Useful for debugging; may leak schema/state in logs. Default: True.
         """
     @property
@@ -627,6 +663,10 @@ class LLParserLimits:
     def precompute_large_lexemes(self) -> bool:
         """Precompute large regexes during lexer construction. Default: True"""
+    @property
+    def verbose_errors(self) -> bool:
+        """Include parser state and grammar in errors. Default: True"""
 def regex_to_lark(regex: str, use_ascii: str = "d") -> str:
     r"""

{llguidance-1.1.1 → llguidance-1.2.0}/python/llguidance/numpy.py RENAMED Viewed

@@ -66,3 +66,17 @@ def fill_next_token_bitmask_par(executor: LLExecutor,
     batch, vocab = bitmask.shape
     assert bitmask.flags["C_CONTIGUOUS"], "Mask must be contiguous"
     executor.unsafe_compute_mask_ptr(matchers, bitmask.ctypes.data, vocab * 4, batch)
+def fill_next_token_bitmask_par_with_draft_tokens(executor: LLExecutor,
+                                matchers: List[Tuple[LLMatcher, int, List[int]]],
+                                bitmask: NDArray[np.int32]) -> None:
+    """
+    Compute the token mask directly into the specified array.
+    For each matcher, provide the index of the target mask.
+    """
+    assert bitmask.dtype == np.int32, "Mask must be int32"
+    assert bitmask.ndim == 2, "Mask must be 2D"
+    batch, vocab = bitmask.shape
+    assert bitmask.flags["C_CONTIGUOUS"], "Mask must be contiguous"
+    executor.unsafe_compute_mask_ptr_with_draft_token(matchers, bitmask.ctypes.data, vocab * 4, batch)

{llguidance-1.1.1 → llguidance-1.2.0}/python/llguidance/torch.py RENAMED Viewed

@@ -66,3 +66,14 @@ def fill_next_token_bitmask_par(executor: LLExecutor,
     assert bitmask.is_contiguous(), "Mask must be contiguous"
     executor.unsafe_compute_mask_ptr(matchers, bitmask.data_ptr(), vocab * 4,
                                      batch)
+def fill_next_token_bitmask_par_with_draft_tokens(executor: LLExecutor,
+                                matchers: List[Tuple[LLMatcher, int, List[int]]],
+                                bitmask: torch.Tensor) -> None:
+    assert bitmask.dtype == torch.int32, "Mask must be int32"
+    assert bitmask.is_cpu, "Mask must be on CPU"
+    assert bitmask.dim() == 2, "Mask must be 2D"
+    batch, vocab = bitmask.shape
+    assert bitmask.is_contiguous(), "Mask must be contiguous"
+    executor.unsafe_compute_mask_ptr_with_draft_token(matchers, bitmask.data_ptr(), vocab * 4, batch)

llguidance 1.1.1__tar.gz → 1.2.0__tar.gz

llguidance 1.1.1tar.gz → 1.2.0tar.gz