llguidance 1.1.1__cp39-abi3-win_amd64.whl → 1.2.0__cp39-abi3-win_amd64.whl

llguidance/_lib.pyi

@@ -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/numpy.py

@@ -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/torch.py

@@ -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.dist-info → llguidance-1.2.0.dist-info}/METADATA

@@ -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.dist-info → llguidance-1.2.0.dist-info}/RECORD

@@ -1,10 +1,10 @@
-llguidance-1.1.1.dist-info/METADATA,sha256=sGRiU_0_1316nEaOxyPtcskGuwk7EUTWa5sTmsUg9Wg,10425
-llguidance-1.1.1.dist-info/WHEEL,sha256=IOvNb_Dp11OuRIpo2OWqTSUXKUtcHN63pkGBkO879SE,94
-llguidance-1.1.1.dist-info/licenses/LICENSE,sha256=mQaUD2Gx8LUz-n2ZuvVReLKAj74RPqUd-_rYVyzNXys,1162
+llguidance-1.2.0.dist-info/METADATA,sha256=nidxJzRXFMnv7QS874s9-I00HSTVvxKFFOunY0O75EY,10436
+llguidance-1.2.0.dist-info/WHEEL,sha256=2XatmAWXBfp_P6DUtFAtbdzzba6f_xbhEtpqsZt_zEg,94
+llguidance-1.2.0.dist-info/licenses/LICENSE,sha256=mQaUD2Gx8LUz-n2ZuvVReLKAj74RPqUd-_rYVyzNXys,1162
 llguidance/__init__.py,sha256=tlrM-GzMdClgSIEblgA4ixP10h51VDJY5ejIP5FFJJw,620
 llguidance/_grammar_from.py,sha256=kEgwhVYBq3NE_O9MGjIRpci7dgOneLC9CvVGWpKEO84,2452
-llguidance/_lib.pyd,sha256=TeIQ9qmb3j3Ov57hPHnmd1BJBVHpYDBGrhUKcUCUwkM,7544832
-llguidance/_lib.pyi,sha256=sxxOUh24-RL6voiR0npIQHj8znMOcrFM43NU1cB8-_c,22870
+llguidance/_lib.pyd,sha256=A4GN_LLSo1y68TyTRyQE-hSRgOXlJrnkRguqyehpYpE,7582208
+llguidance/_lib.pyi,sha256=tc0mb-ZKH3LjodEf_aIdLPXFLuEbqGI3flxQpxZn-Bw,24893
 llguidance/_struct_tag.py,sha256=wEJd9KGZTpS-R1ZXyyRFmr-48aYOZw_tI4PhIRy9h4g,4548
 llguidance/_tokenizer.py,sha256=9jonF41IUypGtV6lVUl9l9FGAyn9HNnDb9HKNDywZuQ,1120
 llguidance/_util.py,sha256=6RMYO61KUynAym7KcbDn9VEMGcGdv7rj2yo6fITC-u8,274
@@ -13,8 +13,8 @@ llguidance/gbnf_to_lark.py,sha256=YsFNyVXrzG3-cjqlcrcuoG1RdquVtZJz43X_lSAXL7Y,17
 llguidance/hf.py,sha256=3LNNqC4_KafPe6bKP-Ba0aGsOVg3MqvuMvY6QYth0fY,1980
 llguidance/llamacpp.py,sha256=2LdNjiNku-8xGcswGEQga57FUsRgzdf4Z4BrLVBUfAg,2352
 llguidance/mlx.py,sha256=Q2cNocnFB3xqCCg614iE0wPcumBFJBvo0pG5JTpShds,2494
-llguidance/numpy.py,sha256=gUQkqJMWWlz9T-Vohtk6pH2XXqP0-RCoaWIOBgRHBhY,2721
+llguidance/numpy.py,sha256=eMIihG5rcPWFMMWjwl-ZPSkcEo8BXKZFUHaIHkxnFhw,3409
 llguidance/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 llguidance/tiktoken.py,sha256=sE4QWj88dANfdM0Y50ZvFI7ku6zBFomUCE4bRTtewF4,1135
-llguidance/torch.py,sha256=SgSutKZzhdlwFPRwD8fxLiPJsD45YYriNbM4Cr21jK0,2863
-llguidance-1.1.1.dist-info/RECORD,,
+llguidance/torch.py,sha256=PzTI7wRiznXnH7u4_iEP6NL2wmL2cN8p-UlR0PE8JJ4,3451
+llguidance-1.2.0.dist-info/RECORD,,

{llguidance-1.1.1.dist-info → llguidance-1.2.0.dist-info}/WHEEL

@@ -1,4 +1,4 @@
 Wheel-Version: 1.0
-Generator: maturin (1.9.1)
+Generator: maturin (1.9.3)
 Root-Is-Purelib: false
 Tag: cp39-abi3-win_amd64