phi-2 token out of bounds error

[nking-1]

The bug
When generating text with phi-2, sometimes the model attempts to sample a token out of bounds of the tokenizer's token list

To Reproduce
This might take multiple attempts due to the random sampling.

from guidance import models
phi2 = models.Transformers("microsoft/phi-2", device_map="mps")

msg_no_e = phi2 + "Hello, " + with_temperature(any_char_but('e'), temperature=0.9)
for i in range(250):
    msg_no_e += with_temperature(any_char_but('e'), temperature=0.9)

msg_no_e

Error trace:

{
	"name": "IndexError",
	"message": "index 51164 is out of bounds for axis 0 with size 50295",
	"stack": "---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
Cell In[12], line 3
      1 msg_no_e = phi2 + \"Hello, \" + with_temperature(any_char_but('e'), temperature=0.9)
      2 for i in range(250):
----> 3     msg_no_e += with_temperature(any_char_but('e'), temperature=0.9)
      5 msg_no_e

File ~/code/ms/guidance/guidance/models/_model.py:915, in Model.__add__(self, value)
    913 # run stateless functions (grammar nodes)
    914 elif isinstance(value, GrammarFunction):
--> 915     out = lm._run_stateless(value)
    917 # run stateful functions
    918 else:
    919     out = value(lm)

File ~/code/ms/guidance/guidance/models/_model.py:1111, in Model._run_stateless(self, stateless_function, temperature, top_p, n)
   1109 delayed_bytes = b\"\"
   1110 # last_is_generated = False
-> 1111 for chunk in gen_obj:
   1112 
   1113     # we make everything full probability if we are not computing uncertainty
   1114     # if not self.engine.compute_log_probs:
   1115     #     chunk.new_bytes_prob = 1.0
   1116     
   1117     # convert the bytes to a string (delaying if we don't yet have a valid unicode string)
   1118     lm.token_count += chunk.new_token_count
   1119     chunk.new_bytes = delayed_bytes + chunk.new_bytes

File ~/code/ms/guidance/guidance/models/_model.py:373, in Engine.__call__(self, parser, grammar, ensure_bos_token)
    371 # loop over the tokens looking for a valid one
    372 for i,sampled_token_ind in enumerate(sampling_order):
--> 373     sampled_token = self.tokenizer.tokens[sampled_token_ind]
    375     # break out if we have reach impossible tokens
    376     if logits[sampled_token_ind] <= -np.inf:

IndexError: index 51164 is out of bounds for axis 0 with size 50295"
}

System info (please complete the following information):

• OS (e.g. Ubuntu, Windows 11, Mac OS, etc.): Mac OS, using mps device
• Guidance Version (guidance.__version__): 0.1.11