remove unneeded torch dependencies

guidance/models/_llama_cpp.py

@@ -9,13 +9,6 @@
 from ._local import Local
 from .._utils import normalize_notebook_stdout_stderr
 
-try:
-    # TODO: can we eliminate the torch requirement for llama.cpp by using numpy in the caller instead?
-    import torch
-    is_torch = True
-except ImportError:
-    is_torch = False
-
 try:
     import llama_cpp
     is_llama_cpp = True
@@ -24,8 +17,6 @@
 
 class LlamaCpp(Local):
     def __init__(self, model=None, tokenizer=None, echo=True, caching=True, temperature=0.0, **kwargs):
-        if not is_torch:
-            raise Exception("Please install PyTorch in order to use guidance.models.LlamaCpp!")
 
         if not is_llama_cpp:
             raise Exception("Please install llama-cpp-python with `pip install llama-cpp-python` in order to use guidance.models.LlamaCpp!")
@@ -139,7 +130,6 @@ def _get_logits(self, token_ids, forced_bytes):
         finally:
             llama_cpp.llama_batch_free(batch)
 
-        logits = torch.from_numpy(logits)
         self._cache_state["logits"] = logits
 
         return logits

guidance/models/_local.py

@@ -1,14 +1,10 @@
-try:
-    import torch
-except ImportError:
-    pass
+import scipy.special
+import scipy.stats
 import numpy as np
 from .._utils import ByteTrie
 from ._model import Model
-# from ..library._string import string
 from .._parser import EarleyCommitParser
-from .._grammar import Terminal, select
-# import numba
+from .._grammar import Terminal
 
 class Local(Model):
     def __init__(self, tokens, bos_token_id, eos_token_id=None, echo=True):
@@ -33,20 +29,7 @@ def _get_logits(self, token_ids, forced_bytes):
         '''A fake method designed to be overriden by subclasses.'''
 
         # pretend to extend the KV cache and update the log probs
-        return torch.randn(len(self.tokens))
-
-    # def _longest_token_match(self, bytes):
-    #     '''Greedy token matching.'''
-    #     trie_pos = self._token_trie
-    #     for i,c in enumerate(bytes):
-    #         if c in trie_pos.children:
-    #             trie_pos = trie_pos.children[c]
-    #         else:
-    #             return bytes[:i], trie_pos.value # note that if there are redudant tokens we choose the one stored in the trie
-    #     if len(trie_pos.children) == 0:
-    #         return bytes[:i+1], trie_pos.value
-    #     else:
-    #         return None,None # more than one token can match these bytes
+        return np.randn(len(self.tokens))
 
     def _tokenize_prefix(self, byte_string):
         '''This is used to speed up the tokenization of long prompts without using the parser.'''
@@ -237,17 +220,17 @@ def __call__(self, grammar, max_tokens=100, n=1, top_p=1, temperature=0.0, ensur
                 # if requested we compute the log probabilities so we can track the probabilities of each node
                 # TODO: we should lower this step to C++ with pybind11
                 if log_probs:
-                    _compute_log_probs(trie, torch.nn.functional.log_softmax(logits, dim=-1).cpu().numpy())
+                    _compute_log_probs(trie, scipy.special.log_softmax(logits, dim=-1))
 
                 # get the sampling order
                 grammar_temp = parser.next_byte_temperature()
                 current_temp = grammar_temp if grammar_temp >= 0 else temperature # we prefer to use the grammar temp when it is specified
                 if current_temp == 0:
-                    sampling_order = torch.argsort(-logits, descending=False).cpu().numpy() # we need numpy so the enumerate below does not get really slow...
+                    sampling_order = np.argsort(-logits) # we need numpy so the enumerate below does not get really slow...
                 else:
                     assert top_p == 1, "Still need to add support for top_p!"
-                    probs = torch.nn.functional.softmax(logits / current_temp, dim=-1)
-                    sampling_order = torch.multinomial(probs, len(probs)).cpu().numpy()
+                    probs = scipy.special.softmax(logits / current_temp, axis=-1)
+                    sampling_order = np.random.choice(len(probs), size=len(probs), p=probs)
 
                 # loop over the tokens looking for a valid one
                 for i,sampled_token_ind in enumerate(sampling_order):

guidance/models/_local_mock.py

@@ -1,15 +1,5 @@
-import re
-import functools
-import time
-import collections
-import os
 import numpy as np
 
-try:
-    import torch
-except ImportError:
-    pass
-
 from ._model import Chat
 from ._local import Local
 
@@ -69,7 +59,7 @@ def _get_logits(self, token_ids, forced_bytes):
                         logits[i] += bias
                     bias /= 2 # if we have multiple matches then they apply with decreasing bias
 
-        return torch.tensor(logits)
+        return logits
 
     def _get_next_tokens(self, byte_string):
         for i,t in enumerate(self.tokens):

guidance/models/_openai.py

@@ -13,14 +13,6 @@
 from ._model import Chat, Instruct
 from ._remote import Remote
 
-
-try:
-    # TODO: can we eliminate the torch requirement for llama.cpp by using numpy in the caller instead?
-    import torch
-    is_torch = True
-except ImportError:
-    is_torch = False
-
 try:
     # TODO: can we eliminate the torch requirement for llama.cpp by using numpy in the caller instead?
     import openai as openai_package

guidance/models/_remote.py

@@ -14,13 +14,6 @@
 from ._local import Local
 
 
-try:
-    # TODO: can we eliminate the torch requirement for llama.cpp by using numpy in the caller instead?
-    import torch
-    is_torch = True
-except ImportError:
-    is_torch = False
-
 # try:
 #     # TODO: can we eliminate the torch requirement for llama.cpp by using numpy in the caller instead?
 #     import vertexai
@@ -190,7 +183,7 @@ def _get_logits(self, token_ids, forced_bytes):
         logits = np.ones(len(self.tokens)) * np.nan
         logits[token_id] = 100
 
-        return torch.tensor(logits) # TODO: the caller need to know to NOT use the 0 entries, but to fail out if that happens
+        return logits
 
     def _get_next_token(self, pos, allow_early_stop=False):
         data = self._shared_state["data"]

guidance/models/transformers/_transformers.py

@@ -124,7 +124,7 @@ def _get_logits(self, token_ids, forced_bytes):
             # save the results
             self._cache_state["past_key_values"] = model_out.past_key_values
             cache_token_ids.extend(new_token_ids)
-            self._cache_state["logits"] = model_out.logits[0, -1, :]
+            self._cache_state["logits"] = model_out.logits[0, -1, :].cpu().numpy()
 
         return self._cache_state["logits"]