Merge pull request #56 from eli5-org/llm-mlx

Support open source LLMs with mlx-lm

Author: lopuhin

docs/source/_notebooks/explain_llm_logprobs.rst

@@ -11,6 +11,9 @@ about its predictions:
 
    LLM token probabilities visualized with eli5.explain_prediction
 
+1. OpenAI models
+----------------
+
 To follow this tutorial you need the ``openai`` library installed and
 working.
 
@@ -64,10 +67,10 @@ properties from a free-form product description:
     json
     {
         "materials": ["metal"],
-        "type": "table lamp",
+        "type": "task lighting",
         "color": "silky matte grey",
         "price": 150.00,
-        "summary": "Stay is a flexible and elegant table lamp designed by Maria Berntsen."
+        "summary": "Stay table lamp with adjustable arm and head for optimal task lighting."
     }
     
 
@@ -311,8 +314,8 @@ We can obtain the original prediction from the explanation object via
 ``explanation.targets[0].target.message.content`` to get the prediction
 text.
 
-Limitations
------------
+2. Limitations
+--------------
 
 Even though above the model confidence matched our expectations, it’s
 not always the case. For example, if we use “Chain of Thought”
@@ -531,6 +534,121 @@ temperatures:
 
 
 
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+
+
+
+
+3. Open Source and other models
+-------------------------------
+
+If an API endpoint can provide ``logprobs`` in the right format, then it
+should work. However few APIs or libraries do provide it, even for open
+source models. One library which is know to work is ``mlx_lm`` (Mac OS
+only), e.g. if you start the server like this:
+
+::
+
+   mlx_lm.server --model mlx-community/Mistral-7B-Instruct-v0.3-4bit
+
+Then you can explain predictions with a custom client:
+
+.. code:: ipython3
+
+    client_custom = openai.OpenAI(base_url="http://localhost:8080/v1", api_key="dummy")
+    eli5.explain_prediction(
+        client_custom,
+        prompt + ' Price should never be zero.',
+        model="mlx-community/Mistral-7B-Instruct-v0.3-4bit",
+    )
+
+
+
+
+.. raw:: html
+
+    
+        <style>
+        table.eli5-weights tr:hover {
+            filter: brightness(85%);
+        }
+    </style>
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+        <p style="margin-bottom: 2.5em; margin-top:0; white-space: pre-wrap;"><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">{
+    </span><span style="background-color: hsl(102.26933456703064, 100.00%, 50.00%)" title="0.969"> </span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> &quot;materials&quot;: [&quot;</span><span style="background-color: hsl(94.45469472360817, 100.00%, 50.00%)" title="0.925">sil</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">ky matte grey metal</span><span style="background-color: hsl(106.55663140845363, 100.00%, 50.00%)" title="0.984">&quot;],</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">
+      &quot;type&quot;:</span><span style="background-color: hsl(67.65625435391846, 100.00%, 50.00%)" title="0.616"> &quot;</span><span style="background-color: hsl(54.98115667185111, 100.00%, 50.00%)" title="0.423">Not</span><span style="background-color: hsl(99.15673841086969, 100.00%, 50.00%)" title="0.954"> specified</span><span style="background-color: hsl(60.184714790030306, 100.00%, 50.00%)" title="0.503"> in</span><span style="background-color: hsl(89.21337460784397, 100.00%, 50.00%)" title="0.882"> the</span><span style="background-color: hsl(78.55493818915068, 100.00%, 50.00%)" title="0.767"> description</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">&quot;,
+      &quot;color&quot;: &quot;</span><span style="background-color: hsl(75.83477075768666, 100.00%, 50.00%)" title="0.732">Not</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> specified in the description&quot;,
+      &quot;price&quot;:</span><span style="background-color: hsl(106.55663140845363, 100.00%, 50.00%)" title="0.984"> </span><span style="background-color: hsl(66.38465137200743, 100.00%, 50.00%)" title="0.597">9</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">9.99,</span><span style="background-color: hsl(70.37163543696023, 100.00%, 50.00%)" title="0.656">
+    </span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">  &quot;summary&quot;: &quot;</span><span style="background-color: hsl(87.74310691905805, 100.00%, 50.00%)" title="0.869">St</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">ay</span><span style="background-color: hsl(96.62538179271525, 100.00%, 50.00%)" title="0.939"> is</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> a</span><span style="background-color: hsl(50.86740763861728, 100.00%, 50.00%)" title="0.362"> flexible</span><span style="background-color: hsl(74.98611970505678, 100.00%, 50.00%)" title="0.720"> and</span><span style="background-color: hsl(53.693017832784676, 100.00%, 50.00%)" title="0.404"> beautiful</span><span style="background-color: hsl(86.3702071513559, 100.00%, 50.00%)" title="0.855"> Dan</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">ish</span><span style="background-color: hsl(78.55493818915068, 100.00%, 50.00%)" title="0.767">-</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">designed</span><span style="background-color: hsl(83.859705332877, 100.00%, 50.00%)" title="0.829"> table</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> lamp</span><span style="background-color: hsl(83.859705332877, 100.00%, 50.00%)" title="0.829"> with</span><span style="background-color: hsl(63.98782501663244, 100.00%, 50.00%)" title="0.561"> a</span><span style="background-color: hsl(57.717412015697704, 100.00%, 50.00%)" title="0.465"> discre</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">et</span><span style="background-color: hsl(102.26933456703064, 100.00%, 50.00%)" title="0.969"> switch</span><span style="background-color: hsl(73.36326933304912, 100.00%, 50.00%)" title="0.698"> and</span><span style="background-color: hsl(52.049284214496396, 100.00%, 50.00%)" title="0.380"> adjust</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">able</span><span style="background-color: hsl(68.98323009885026, 100.00%, 50.00%)" title="0.636"> arm</span><span style="background-color: hsl(90.80119410884282, 100.00%, 50.00%)" title="0.896"> and</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> head</span><span style="background-color: hsl(78.55493818915068, 100.00%, 50.00%)" title="0.767">,</span><span style="background-color: hsl(65.7679832651431, 100.00%, 50.00%)" title="0.588"> ideal</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> for</span><span style="background-color: hsl(69.66933255201431, 100.00%, 50.00%)" title="0.646"> office</span><span style="background-color: hsl(106.55663140845363, 100.00%, 50.00%)" title="0.984"> task</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000"> lighting</span><span style="background-color: hsl(102.26933456703064, 100.00%, 50.00%)" title="0.969">.&quot;</span><span style="background-color: hsl(120.0, 100.00%, 50.00%)" title="1.000">
+    }</span></p>
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
+    
 
 
 

eli5/llm/explain_prediction.py

@@ -1,8 +1,10 @@
 import math
-from typing import Union
+import warnings
+from typing import Optional, Union
 
 import openai
-from openai.types.chat.chat_completion import ChoiceLogprobs, ChatCompletion
+from openai.types.chat.chat_completion import (
+    ChatCompletion, ChatCompletionTokenLogprob, ChoiceLogprobs)
 
 from eli5.base import Explanation, TargetExplanation, WeightedSpans, DocWeightedSpans
 from eli5.explain import explain_prediction
@@ -49,15 +51,15 @@ def explain_prediction_openai_logprobs(logprobs: ChoiceLogprobs, doc=None):
 
 @explain_prediction.register(ChatCompletion)
 def explain_prediction_openai_completion(
-        chat_completion: ChatCompletion, doc=None):
+        completion: ChatCompletion, doc=None):
     """ Creates an explanation of the ChatCompletion's logprobs
     highlighting them proportionally to the log probability.
     More likely tokens are highlighted in green,
     while unlikely tokens are highlighted in red.
     ``doc`` argument is ignored.
     """
     targets = []
-    for choice in chat_completion.choices:
+    for choice in completion.choices:
         if choice.logprobs is None:
             raise ValueError('Predictions must be obtained with logprobs enabled')
         target, = explain_prediction_openai_logprobs(choice.logprobs).targets
@@ -92,8 +94,54 @@ def explain_prediction_openai_client(
     else:
         messages = doc
     kwargs['logprobs'] = True
-    chat_completion = client.chat.completions.create(
+    completion = client.chat.completions.create(
         messages=messages,  # type: ignore
         model=model,
         **kwargs)
-    return explain_prediction_openai_completion(chat_completion)
+    for choice in completion.choices:
+        _recover_logprobs(choice.logprobs, model)
+        if choice.logprobs is None:
+            raise ValueError('logprobs not found, likely API does not support them')
+        if choice.logprobs.content is None:
+            raise ValueError(f'logprobs.content is empty: {choice.logprobs}')
+    return explain_prediction_openai_completion(completion)
+
+
+def _recover_logprobs(logprobs: Optional[ChoiceLogprobs], model: str):
+    """ Some servers don't populate logprobs.content, try to recover it.
+    """
+    if logprobs is None:
+        return
+    if logprobs.content is not None:
+        return
+    if not (
+            getattr(logprobs, 'token_logprobs', None) and
+            getattr(logprobs, 'tokens', None)):
+        return
+    assert hasattr(logprobs, 'token_logprobs')  # for mypy
+    assert hasattr(logprobs, 'tokens')  # for mypy
+    try:
+        import tokenizers
+    except ImportError:
+        warnings.warn('tokenizers library required to recover logprobs.content')
+        return
+    try:
+        tokenizer = tokenizers.Tokenizer.from_pretrained(model)
+    except Exception:
+        warnings.warn(f'could not load tokenizer for {model} with tokenizers library')
+        return
+    assert len(logprobs.token_logprobs) == len(logprobs.tokens)
+    # get tokens as strings with spaces, is there any better way?
+    text = tokenizer.decode(logprobs.tokens)
+    encoded = tokenizer.encode(text, add_special_tokens=False)
+    text_tokens = [text[start:end] for (start, end) in encoded.offsets]
+    logprobs.content = []
+    for logprob, token in zip(logprobs.token_logprobs, text_tokens):
+        logprobs.content.append(
+            ChatCompletionTokenLogprob(
+                token=token,
+                bytes=list(map(int, token.encode('utf8'))),
+                logprob=logprob,
+                top_logprobs=[],  # we could recover that too
+            )
+        )

notebooks/explain_llm_logprobs.ipynb

@@ -3,6 +3,7 @@
 from unittest.mock import Mock
 
 pytest.importorskip('openai')
+pytest.importorskip('transformers')
 from openai.types.chat.chat_completion import (
     ChoiceLogprobs,
     ChatCompletion,
@@ -11,6 +12,7 @@
     Choice,
 )
 from openai import Client
+import transformers
 
 import eli5
 from eli5.base import Explanation
@@ -40,20 +42,28 @@ def example_logprobs():
 
 @pytest.fixture
 def example_completion(example_logprobs):
+    return create_completion(
+        model='gpt-4o-2024-08-06',
+        logprobs=example_logprobs,
+        message=ChatCompletionMessage(
+            content=''.join(x.token for x in example_logprobs.content),
+            role='assistant',
+        ),
+    )
+
+
+def create_completion(model, logprobs, message):
     return ChatCompletion(
         id='chatcmpl-x',
         created=1743590849,
-        model='gpt-4o-2024-08-06',
+        model=model,
         object='chat.completion',
         choices=[
             Choice(
-                logprobs=example_logprobs,
+                logprobs=logprobs,
                 finish_reason='stop',
                 index=0,
-                message=ChatCompletionMessage(
-                    content=''.join(x.token for x in example_logprobs.content),
-                    role='assistant',
-                ),
+                message=message,
             )
         ],
     )
@@ -100,7 +110,35 @@ def __init__(self, chat_return_value):
 def test_explain_prediction_openai_client(monkeypatch, example_completion):
     client = MockClient(example_completion)
 
-    explanation = eli5.explain_prediction(client, doc="Hello world", model="gpt-4o")
+    explanation = eli5.explain_prediction(client, doc="Hello world world", model="gpt-4o")
+    _assert_explanation_structure_and_html(explanation)
+
+    client.chat.completions.create.assert_called_once()
+
+
+def test_explain_prediction_openai_client_mlx(monkeypatch):
+    model = "mlx-community/Mistral-7B-Instruct-v0.3-4bit"
+    tokenizer = transformers.AutoTokenizer.from_pretrained(model)
+
+    text = 'Hello world world'
+    tokens = tokenizer.encode(text, add_special_tokens=False)
+    assert len(tokens) == 3
+    logprobs = ChoiceLogprobs(
+        token_logprobs=[
+            math.log(0.9),
+            math.log(0.2),
+            math.log(0.4),
+        ],
+        tokens=tokens,
+    )
+    completion = create_completion(
+        model=model,
+        logprobs=logprobs,
+        message=ChatCompletionMessage(content=text, role='assistant'),
+    )
+    client = MockClient(completion)
+
+    explanation = eli5.explain_prediction(client, doc=text, model=model)
     _assert_explanation_structure_and_html(explanation)
 
     client.chat.completions.create.assert_called_once()

tests/test_llm_explain_prediction.py

@@ -30,6 +30,7 @@ deps=
     pandas
     sklearn-crfsuite
     openai
+    tokenizers
 commands=
     pip install -e .
     py.test --doctest-modules \