lemonade-sdk 7.0.4__py3-none-any.whl → 8.0.0__py3-none-any.whl

lemonade/tools/huggingface/load.py

@@ -0,0 +1,235 @@
+import argparse
+from typing import Dict, Optional
+import json
+from lemonade.tools import FirstTool
+from lemonade.state import State
+import lemonade.common.status as status
+import lemonade.common.printing as printing
+from lemonade.cache import Keys
+
+
+class HuggingfaceLoad(FirstTool):
+    """
+    Load an LLM as a torch.nn.Module using the Hugging Face transformers
+    from_pretrained() API.
+
+    Expected input: a checkpoint to load
+
+    Output state produced:
+        - state.model: instance of torch.nn.Module that implements an LLM.
+        - state.inputs: tokenized example inputs to the model, in the form of a
+            dictionary of kwargs.
+        - state.tokenizer: instance of Hugging Face PretrainedTokenizer.
+        - state.dtype: data type of the model.
+        - state.checkpoint: pretrained checkpoint used to load the model.
+    """
+
+    unique_name = "huggingface-load"
+
+    def _imports(self):
+        pass
+
+    def __init__(self):
+        super().__init__(monitor_message="Loading Huggingface checkpoint")
+
+        self.status_stats = [Keys.DTYPE]
+
+    @staticmethod
+    def parser(add_help: bool = True) -> argparse.ArgumentParser:
+        parser = __class__.helpful_parser(
+            short_description="Load an LLM in PyTorch using huggingface transformers",
+            add_help=add_help,
+        )
+
+        default_dtype = "float32"
+        parser.add_argument(
+            "--dtype",
+            "-d",
+            required=False,
+            default=default_dtype,
+            help=f"Data type to load the model in (default: {default_dtype}).",
+        )
+
+        choices = ["cpu", "cuda"]
+        for cuda in range(15):
+            choices.append(f"cuda:{cuda}")
+        parser.add_argument(
+            "--device",
+            required=False,
+            default=None,
+            choices=choices,
+            help="Move the model and inputs to a device using the .to() method "
+            "(default: don't call the .to() method)",
+        )
+
+        parser.add_argument(
+            "--load-kwargs",
+            required=False,
+            default="{}",
+            type=json.loads,
+            help="Arbitrary kwargs, in json format, that will be passed as "
+            "from_pretrained(**kwargs). "
+            r"Example: --load-kwargs='{\"trust_remote_code\": true} would result in "
+            "from_pretrained(trust_remote_code=True)",
+        )
+
+        parser.add_argument(
+            "--channels-last",
+            default=True,
+            type=bool,
+            help="Whether to format the model in memory using "
+            "channels-last (default: True)",
+        )
+
+        return parser
+
+    def parse(self, state: State, args, known_only=True) -> argparse.Namespace:
+
+        from lemonade.tools.huggingface.utils import str_to_dtype
+
+        parsed_args = super().parse(state, args, known_only)
+
+        # Save stats about the user's input (do this prior to decoding)
+        state.save_stat(Keys.CHECKPOINT, parsed_args.input)
+        state.save_stat(Keys.DTYPE, parsed_args.dtype)
+
+        # Decode dtype arg into a torch value
+        parsed_args.dtype = str_to_dtype[parsed_args.dtype]
+
+        return parsed_args
+
+    def run(
+        self,
+        state: State,
+        input: str = "",
+        dtype: "torch.dtype" = None,
+        device: Optional[str] = None,
+        load_kwargs: Optional[Dict] = None,
+        channels_last: bool = True,
+    ) -> State:
+        # Import expensive modules at runtime
+        import transformers
+        import torch
+
+        from lemonade.tools.huggingface.utils import (
+            HuggingfaceTokenizerAdapter,
+            HuggingfaceAdapter,
+        )
+        from lemonade.common.network import (
+            is_offline,
+            get_base_model,
+        )
+
+        # Set default dtype
+        if dtype is None:
+            dtype_to_use = torch.float32
+        else:
+            dtype_to_use = dtype
+
+        # Auto-detect offline status
+        offline = is_offline()
+        if offline:
+            printing.log_warning(
+                "Network connectivity to huggingface.co not detected. Running in offline mode."
+            )
+
+        checkpoint = input
+
+        if load_kwargs is None:
+            load_kwargs_to_use = {}
+        else:
+            load_kwargs_to_use = load_kwargs
+
+        # Add local_files_only to kwargs in offline mode
+        if offline:
+            load_kwargs_to_use["local_files_only"] = True
+
+        if vars(state).get(Keys.MODEL):
+            raise ValueError("HuggingfaceLoad must be the first tool in the sequence")
+
+        try:
+            model = transformers.AutoModelForCausalLM.from_pretrained(
+                checkpoint,
+                torch_dtype=dtype_to_use,
+                low_cpu_mem_usage=True,
+                **load_kwargs_to_use,
+            )
+        except Exception as e:
+            if offline and "Can't load config for" in str(e):
+                raise ValueError(
+                    f"Cannot load model {checkpoint} in offline mode. "
+                    f"The model files may not be available locally. Original error: {str(e)}"
+                )
+            raise
+
+        # Only call the model.to() method if an argument to this function
+        # provides a reason to do so
+        to_args = {}
+        if channels_last:
+            to_args["memory_format"] = torch.channels_last
+        if device:
+            to_args["device"] = device
+        if to_args:
+            model.to(**to_args)
+
+        model = model.eval()
+
+        try:
+            tokenizer_kwargs = {
+                "use_fast": False,
+                "model_max_length": 4096,
+                "padding_side": "left",
+            }
+            if offline:
+                tokenizer_kwargs["local_files_only"] = True
+
+            tokenizer = transformers.AutoTokenizer.from_pretrained(
+                checkpoint, **tokenizer_kwargs
+            )
+        except ValueError as e:
+            # Sometimes those specific tokenizer flags are not supported, in which
+            # case we try to just load a simple tokenizer
+            tokenizer_kwargs = {}
+            if offline:
+                tokenizer_kwargs["local_files_only"] = True
+
+            try:
+                tokenizer = transformers.AutoTokenizer.from_pretrained(
+                    checkpoint, **tokenizer_kwargs
+                )
+            except Exception as e:
+                if offline and "Can't load tokenizer for" in str(e):
+                    raise ValueError(
+                        f"Cannot load tokenizer for {checkpoint} in offline mode. "
+                        f"The tokenizer files may not be available locally. "
+                        f"Original error: {str(e)}"
+                    )
+                raise
+
+        # Pass the model and inputs into state
+        state.model = HuggingfaceAdapter(model, dtype_to_use, device, tokenizer)
+
+        state.tokenizer = HuggingfaceTokenizerAdapter(tokenizer, device)
+        state.dtype = dtype_to_use
+        state.checkpoint = checkpoint
+        state.device = device
+
+        # Save stats about the model
+        state.save_stat(Keys.CHECKPOINT, checkpoint)
+        state.save_stat(Keys.DTYPE, str(dtype_to_use).split(".")[1])
+        state.save_stat(Keys.DEVICE, device)
+
+        # Get base model information
+        base_model = get_base_model(checkpoint)
+        if base_model is not None:
+            state.save_stat("base_model", base_model)
+
+        # Create a UniqueInvocationInfo and ModelInfo so that we can display status
+        # at the end of the sequence
+        status.add_to_state(state=state, name=input, model=model)
+
+        return state
+
+
+# This file was originally licensed under Apache 2.0. It has been modified.
+# Modifications Copyright (c) 2025 AMD

lemonade/tools/{huggingface_load.py → huggingface/utils.py}

@@ -1,16 +1,12 @@
-import argparse
-from typing import Dict, Optional
-import json
-import socket
+from typing import Dict, List, Tuple
+import time
+from contextlib import nullcontext
 import transformers
 import torch
-from huggingface_hub import model_info
 from lemonade.state import State
-import lemonade.common.status as status
-import lemonade.common.printing as printing
-from lemonade.tools import FirstTool
-from lemonade.tools.adapter import ModelAdapter, TokenizerAdapter
-from lemonade.cache import Keys
+from lemonade.tools.adapter import TokenizerAdapter
+from lemonade.tools.adapter import ModelAdapter
+from lemonade.tools.bench import Bench
 
 # Command line interfaces for tools will use string inputs for data
 # types, however the internal tool logic will need to know the actual
@@ -62,249 +58,6 @@ class HuggingfaceTokenizerAdapter(TokenizerAdapter):
         return self.tokenizer.save_pretrained(model_dir, **kwargs)
 
 
-def is_offline():
-    """
-    Check if the system is offline by attempting to connect to huggingface.co.
-
-    Returns:
-        bool: True if the system is offline (cannot connect to huggingface.co),
-              False otherwise.
-    """
-    try:
-        socket.gethostbyname("huggingface.co")
-        return False
-    except socket.gaierror:
-        return True
-
-
-def get_base_model(checkpoint: str) -> Optional[str]:
-    """
-    Get the base model information for a given checkpoint from the Hugging Face Hub.
-    Will auto-detect if we're offline and skip the network call in that case.
-
-    Args:
-        checkpoint: The model checkpoint to query
-
-    Returns:
-        The base model name if found, or None if not found or error occurs
-    """
-    # Skip network call in offline mode
-    if is_offline():
-        return None
-
-    try:
-        info = model_info(checkpoint)
-        if info.cardData and "base_model" in info.cardData:
-            if info.cardData["base_model"] is not None:
-                # This is a derived model
-                return info.cardData["base_model"]
-            else:
-                # This is itself a base model
-                return [checkpoint]
-    except Exception:  # pylint: disable=broad-except
-        pass
-    return None
-
-
-class HuggingfaceLoad(FirstTool):
-    """
-    Load an LLM as a torch.nn.Module using the Hugging Face transformers
-    from_pretrained() API.
-
-    Expected input: a checkpoint to load
-
-    Output state produced:
-        - state.model: instance of torch.nn.Module that implements an LLM.
-        - state.inputs: tokenized example inputs to the model, in the form of a
-            dictionary of kwargs.
-        - state.tokenizer: instance of Hugging Face PretrainedTokenizer.
-        - state.dtype: data type of the model.
-        - state.checkpoint: pretrained checkpoint used to load the model.
-    """
-
-    unique_name = "huggingface-load"
-
-    def __init__(self):
-        super().__init__(monitor_message="Loading Huggingface checkpoint")
-
-        self.status_stats = [Keys.DTYPE]
-
-    @staticmethod
-    def parser(add_help: bool = True) -> argparse.ArgumentParser:
-        parser = __class__.helpful_parser(
-            short_description="Load an LLM in PyTorch using huggingface transformers",
-            add_help=add_help,
-        )
-
-        default_dtype = "float32"
-        parser.add_argument(
-            "--dtype",
-            "-d",
-            required=False,
-            default=default_dtype,
-            help=f"Data type to load the model in (default: {default_dtype}).",
-        )
-
-        choices = ["cpu", "cuda"]
-        for cuda in range(15):
-            choices.append(f"cuda:{cuda}")
-        parser.add_argument(
-            "--device",
-            required=False,
-            default=None,
-            choices=choices,
-            help="Move the model and inputs to a device using the .to() method "
-            "(default: don't call the .to() method)",
-        )
-
-        parser.add_argument(
-            "--load-kwargs",
-            required=False,
-            default="{}",
-            type=json.loads,
-            help="Arbitrary kwargs, in json format, that will be passed as "
-            "from_pretrained(**kwargs). "
-            r"Example: --load-kwargs='{\"trust_remote_code\": true} would result in "
-            "from_pretrained(trust_remote_code=True)",
-        )
-
-        parser.add_argument(
-            "--channels-last",
-            default=True,
-            type=bool,
-            help="Whether to format the model in memory using "
-            "channels-last (default: True)",
-        )
-
-        return parser
-
-    def parse(self, state: State, args, known_only=True) -> argparse.Namespace:
-
-        parsed_args = super().parse(state, args, known_only)
-
-        # Save stats about the user's input (do this prior to decoding)
-        state.save_stat(Keys.CHECKPOINT, parsed_args.input)
-        state.save_stat(Keys.DTYPE, parsed_args.dtype)
-
-        # Decode dtype arg into a torch value
-        parsed_args.dtype = str_to_dtype[parsed_args.dtype]
-
-        return parsed_args
-
-    def run(
-        self,
-        state: State,
-        input: str = "",
-        dtype: torch.dtype = torch.float32,
-        device: Optional[str] = None,
-        load_kwargs: Optional[Dict] = None,
-        channels_last: bool = True,
-    ) -> State:
-        # Auto-detect offline status
-        offline = is_offline()
-        if offline:
-            printing.log_warning(
-                "Network connectivity to huggingface.co not detected. Running in offline mode."
-            )
-
-        checkpoint = input
-
-        if load_kwargs is None:
-            load_kwargs_to_use = {}
-        else:
-            load_kwargs_to_use = load_kwargs
-
-        # Add local_files_only to kwargs in offline mode
-        if offline:
-            load_kwargs_to_use["local_files_only"] = True
-
-        if vars(state).get(Keys.MODEL):
-            raise ValueError("HuggingfaceLoad must be the first tool in the sequence")
-
-        try:
-            model = transformers.AutoModelForCausalLM.from_pretrained(
-                checkpoint,
-                torch_dtype=dtype,
-                low_cpu_mem_usage=True,
-                **load_kwargs_to_use,
-            )
-        except Exception as e:
-            if offline and "Can't load config for" in str(e):
-                raise ValueError(
-                    f"Cannot load model {checkpoint} in offline mode. "
-                    f"The model files may not be available locally. Original error: {str(e)}"
-                )
-            raise
-
-        # Only call the model.to() method if an argument to this function
-        # provides a reason to do so
-        to_args = {}
-        if channels_last:
-            to_args["memory_format"] = torch.channels_last
-        if device:
-            to_args["device"] = device
-        if to_args:
-            model.to(**to_args)
-
-        model = model.eval()
-
-        try:
-            tokenizer_kwargs = {
-                "use_fast": False,
-                "model_max_length": 4096,
-                "padding_side": "left",
-            }
-            if offline:
-                tokenizer_kwargs["local_files_only"] = True
-
-            tokenizer = transformers.AutoTokenizer.from_pretrained(
-                checkpoint, **tokenizer_kwargs
-            )
-        except ValueError as e:
-            # Sometimes those specific tokenizer flags are not supported, in which
-            # case we try to just load a simple tokenizer
-            tokenizer_kwargs = {}
-            if offline:
-                tokenizer_kwargs["local_files_only"] = True
-
-            try:
-                tokenizer = transformers.AutoTokenizer.from_pretrained(
-                    checkpoint, **tokenizer_kwargs
-                )
-            except Exception as e:
-                if offline and "Can't load tokenizer for" in str(e):
-                    raise ValueError(
-                        f"Cannot load tokenizer for {checkpoint} in offline mode. "
-                        f"The tokenizer files may not be available locally. "
-                        f"Original error: {str(e)}"
-                    )
-                raise
-
-        # Pass the model and inputs into state
-        state.model = HuggingfaceAdapter(model, dtype, device, tokenizer)
-
-        state.tokenizer = HuggingfaceTokenizerAdapter(tokenizer, device)
-        state.dtype = dtype
-        state.checkpoint = checkpoint
-        state.device = device
-
-        # Save stats about the model
-        state.save_stat(Keys.CHECKPOINT, checkpoint)
-        state.save_stat(Keys.DTYPE, str(dtype).split(".")[1])
-        state.save_stat(Keys.DEVICE, device)
-
-        # Get base model information
-        base_model = get_base_model(checkpoint)
-        if base_model is not None:
-            state.save_stat("base_model", base_model)
-
-        # Create a UniqueInvocationInfo and ModelInfo so that we can display status
-        # at the end of the sequence
-        status.add_to_state(state=state, name=input, model=model)
-
-        return state
-
-
 class HuggingfaceAdapter(ModelAdapter):
     """
     Wrapper class for Huggingface LLMs that handle generation arguments
@@ -522,5 +275,84 @@ class HuggingfaceAdapter(ModelAdapter):
         return text_offset, token_log_probs, token_strings, top_logprobs_list
 
 
-# This file was originally licensed under Apache 2.0. It has been modified.
-# Modifications Copyright (c) 2025 AMD
+def benchmark_huggingface_llm(
+    model: torch.nn.Module,
+    tokenizer,
+    input_ids,
+    dtype,
+    num_beams: int,
+    target_output_tokens: int,
+    iterations: int,
+    warmup_iterations: int,
+    report_progress_fn,
+) -> List[Tuple[float, int]]:
+
+    amp_enabled = True if (dtype == torch.float16 or dtype == torch.bfloat16) else False
+    # The "if amp_enabled else nullcontext()" is to get around a bug in PyTorch 2.1
+    # where torch.cpu.amp.autocast(enabled=False) does nothing
+    with (
+        torch.cpu.amp.autocast(enabled=amp_enabled, dtype=dtype)
+        if amp_enabled
+        else nullcontext()
+    ):
+
+        per_iteration_result = []
+        tokens_out_len_list = []
+
+        # Early stopping is only a valid parameter with multiple beams
+        early_stopping = num_beams > 1
+
+        with torch.no_grad(), torch.inference_mode():
+            # Don't capture time for warmup
+            for count in range(warmup_iterations):
+                outputs = model.generate(
+                    input_ids,
+                    num_beams=num_beams,
+                    max_new_tokens=target_output_tokens,
+                    min_new_tokens=target_output_tokens,
+                    early_stopping=early_stopping,
+                    pad_token_id=tokenizer.eos_token_id,
+                )
+                tokens_out_len_list.append(outputs.shape[1] - input_ids.shape[1])
+                report_progress_fn((count + 1) / (warmup_iterations + iterations))
+
+            for count in range(iterations):
+                # CUDA synchronization is required prior to GPU benchmarking
+                # This has no negative effect on CPU-only benchmarks, and is more robust than
+                # checking `model.device == "cuda"` since it applies to multi-GPU environments
+                # Synchronization is done before collecting the start time because this will
+                # ensure that the GPU has finished initialization tasks such as loading weights
+                if torch.cuda.is_available():
+                    torch.cuda.synchronize()
+                start_time = time.perf_counter()
+
+                outputs = model.generate(
+                    input_ids,
+                    num_beams=num_beams,
+                    max_new_tokens=target_output_tokens,
+                    min_new_tokens=target_output_tokens,
+                    early_stopping=early_stopping,
+                    pad_token_id=tokenizer.eos_token_id,
+                )
+
+                if torch.cuda.is_available():
+                    torch.cuda.synchronize()
+                end_time = time.perf_counter()
+
+                latency = end_time - start_time
+
+                token_len = outputs.shape[1] - input_ids.shape[1]
+                tokens_out_len_list.append(token_len)
+
+                # Only count an iteration if it produced enough tokens
+                if token_len >= target_output_tokens:
+                    per_iteration_result.append((latency, token_len))
+
+                report_progress_fn(
+                    (warmup_iterations + count + 1) / (warmup_iterations + iterations)
+                )
+
+        if not per_iteration_result:
+            raise Bench.not_enough_tokens(target_output_tokens)
+
+    return per_iteration_result, tokens_out_len_list

lemonade/tools/humaneval.py

@@ -2,9 +2,7 @@ import argparse
 import os
 import csv
 from typing import Dict, Optional, Any
-import requests
-from human_eval.data import write_jsonl, read_problems
-from human_eval.evaluation import evaluate_functional_correctness
+
 
 from lemonade.state import State
 from lemonade.tools import Tool
@@ -95,6 +93,7 @@ class AccuracyHumaneval(Tool):
         Returns:
             Updated state with evaluation results
         """
+
         # Validate required state components
         if not hasattr(state, "model") or not hasattr(state, "tokenizer"):
             raise ValueError("State must contain both 'model' and 'tokenizer'")
@@ -128,6 +127,9 @@ class AccuracyHumaneval(Tool):
 
     def _download_dataset(self, output_path: str) -> None:
         """Download HumanEval dataset if not already present."""
+
+        import requests
+
         if os.path.exists(output_path):
             printing.log_info(f"Dataset already exists at: {output_path}")
             return
@@ -170,6 +172,10 @@ class AccuracyHumaneval(Tool):
         Returns:
             Dictionary containing evaluation metrics
         """
+
+        from human_eval.data import write_jsonl, read_problems
+        from human_eval.evaluation import evaluate_functional_correctness
+
         dataset = read_problems(data_path)
 
         # Limit to first N problems

lemonade/tools/{llamacpp_bench.py → llamacpp/bench.py}

@@ -3,7 +3,7 @@ import statistics
 from statistics import StatisticsError
 from lemonade.state import State
 from lemonade.cache import Keys
-from lemonade.tools.llamacpp import LlamaCppAdapter
+from lemonade.tools.llamacpp.load import LlamaCppAdapter
 from lemonade.tools.bench import Bench