lemonade-sdk 8.1.11__tar.gz → 8.1.12__tar.gz

{lemonade_sdk-8.1.11/src/lemonade_sdk.egg-info → lemonade_sdk-8.1.12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lemonade-sdk
-Version: 8.1.11
+Version: 8.1.12
 Summary: Lemonade SDK: Your LLM Aide for Validation and Deployment
 Author-email: lemonade@amd.com
 Requires-Python: >=3.10, <3.14
@@ -264,6 +264,7 @@ This project is:
   - [OnnxRuntime GenAI](https://github.com/microsoft/onnxruntime-genai)
   - [Hugging Face Hub](https://github.com/huggingface/huggingface_hub)
   - [OpenAI API](https://github.com/openai/openai-python)
+  - [IRON/MLIR-AIE](https://github.com/Xilinx/mlir-aie)
   - and more...
 - Accelerated by mentorship from the OCV Catalyst program.
 - Licensed under the [Apache 2.0 License](https://github.com/lemonade-sdk/lemonade/blob/main/LICENSE).

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/README.md

@@ -207,6 +207,7 @@ This project is:
   - [OnnxRuntime GenAI](https://github.com/microsoft/onnxruntime-genai)
   - [Hugging Face Hub](https://github.com/huggingface/huggingface_hub)
   - [OpenAI API](https://github.com/openai/openai-python)
+  - [IRON/MLIR-AIE](https://github.com/Xilinx/mlir-aie)
   - and more...
 - Accelerated by mentorship from the OCV Catalyst program.
 - Licensed under the [Apache 2.0 License](https://github.com/lemonade-sdk/lemonade/blob/main/LICENSE).

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/src/lemonade/cache.py

@@ -43,7 +43,11 @@ def build_name(input_name):
     """
 
     if os.path.isdir(input_name):
+        # Input is a folder so no good way to determine a model name
         input_name_sanitized = "local_model"
+    elif os.path.isfile(input_name):
+        # Use the filename without its extension
+        input_name_sanitized = os.path.splitext(os.path.basename(input_name))[0]
     else:
         # Sanitize the input name
         input_name_sanitized = input_name.replace("/", "_")
@@ -63,8 +67,9 @@ class Keys:
     TOKEN_GENERATION_TOKENS_PER_SECOND = "token_generation_tokens_per_second"
     STD_DEV_TOKENS_PER_SECOND = "std_dev_tokens_per_second"
     SECONDS_TO_FIRST_TOKEN = "seconds_to_first_token"
-    PREFILL_TOKENS_PER_SECOND = "prefill_tokens_per_second"
     STD_DEV_SECONDS_TO_FIRST_TOKEN = "std_dev_seconds_to_first_token"
+    PREFILL_TOKENS_PER_SECOND = "prefill_tokens_per_second"
+    STD_DEV_PREFILL_TOKENS_PER_SECOND = "std_dev_prefill_tokens_per_second"
     CHECKPOINT = "checkpoint"
     DTYPE = "dtype"
     PROMPT = "prompt"

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/src/lemonade/common/status.py

@@ -112,10 +112,10 @@ class UniqueInvocationInfo(BasicInfo):
         if print_file_name:
             print(f"{self.script_name}{self.extension}:")
 
-        # Print invocation about the model (only applies to scripts, not ONNX files or
+        # Print invocation about the model (only applies to scripts, not ONNX or GGUF files, nor
         # LLMs, which have no extension)
         if not (
-            self.extension == ".onnx"
+            self.extension in [".onnx", ".gguf"]
             or self.extension == build.state_file_name
             or self.extension == ""
         ):
@@ -138,7 +138,7 @@ class UniqueInvocationInfo(BasicInfo):
 
         if self.depth == 0:
             print(f"{self.indent}\tLocation:\t{self.file}", end="")
-            if self.extension == ".onnx":
+            if self.extension in [".onnx", ".gguf"]:
                 print()
             else:
                 print(f", line {self.line}")
@@ -314,7 +314,7 @@ class UniqueInvocationInfo(BasicInfo):
         Print information about a given model or submodel.
         """
 
-        if self.extension == ".onnx" or self.extension == "":
+        if self.extension in [".onnx", ".gguf"] or self.extension == "":
             self.indent = "\t" * (2 * self.depth)
         else:
             self.indent = "\t" * (2 * self.depth + 1)

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/src/lemonade/tools/bench.py

@@ -29,7 +29,9 @@ class Bench(Tool, ABC):
             Keys.SECONDS_TO_FIRST_TOKEN,
             Keys.STD_DEV_SECONDS_TO_FIRST_TOKEN,
             Keys.TOKEN_GENERATION_TOKENS_PER_SECOND,
+            Keys.STD_DEV_TOKENS_PER_SECOND,
             Keys.PREFILL_TOKENS_PER_SECOND,
+            Keys.STD_DEV_PREFILL_TOKENS_PER_SECOND,
             Keys.PROMPT_TOKENS,
             Keys.RESPONSE_TOKENS,
             Keys.MAX_MEMORY_USED_GBYTE,
@@ -42,7 +44,9 @@ class Bench(Tool, ABC):
         self.mean_time_to_first_token_list = []
         self.std_dev_time_to_first_token_list = []
         self.prefill_tokens_per_second_list = []
+        self.std_dev_prefill_tokens_per_second_list = []
         self.token_generation_tokens_per_second_list = []
+        self.std_dev_token_generation_tokens_per_second_list = []
         self.max_memory_used_gb_list = []
 
         # Max memory used can only be measured on Windows systems
@@ -88,7 +92,7 @@ class Bench(Tool, ABC):
             default=[str(default_prompt_length)],
             metavar="PROMPT",
             help="Input one or more prompts to the LLM. Three formats are supported. "
-            "1) integer: use a synthetic prompt with the specified length "
+            "1) integer: use a synthetic prompt with the specified token length "
             "2) str: use a user-provided prompt string "
             "3) path/to/prompt.txt: load the prompt from a text file. "
             f"(default: {default_prompt_length}) ",
@@ -246,10 +250,27 @@ class Bench(Tool, ABC):
             Keys.PREFILL_TOKENS_PER_SECOND,
             self.get_item_or_list(self.prefill_tokens_per_second_list),
         )
+        if not all(
+            element is None for element in self.std_dev_prefill_tokens_per_second_list
+        ):
+            state.save_stat(
+                Keys.STD_DEV_PREFILL_TOKENS_PER_SECOND,
+                self.get_item_or_list(self.std_dev_prefill_tokens_per_second_list),
+            )
         state.save_stat(
             Keys.TOKEN_GENERATION_TOKENS_PER_SECOND,
             self.get_item_or_list(self.token_generation_tokens_per_second_list),
         )
+        if not all(
+            element is None
+            for element in self.std_dev_token_generation_tokens_per_second_list
+        ):
+            state.save_stat(
+                Keys.STD_DEV_TOKENS_PER_SECOND,
+                self.get_item_or_list(
+                    self.std_dev_token_generation_tokens_per_second_list
+                ),
+            )
         if self.save_max_memory_used:
             state.save_stat(
                 Keys.MAX_MEMORY_USED_GBYTE,

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/src/lemonade/tools/flm/utils.py

@@ -13,7 +13,7 @@ import requests
 from packaging.version import Version
 
 
-FLM_MINIMUM_VERSION = "0.9.10"
+FLM_MINIMUM_VERSION = "0.9.12"
 
 
 def check_flm_version() -> Optional[str]:

lemonade_sdk-8.1.12/src/lemonade/tools/llamacpp/bench.py

@@ -0,0 +1,224 @@
+import argparse
+import statistics
+from statistics import StatisticsError
+from lemonade.state import State
+from lemonade.tools.tool import Tool
+from lemonade.tools.llamacpp.utils import LlamaCppAdapter
+from lemonade.tools.bench import (
+    Bench,
+    default_prompt_length,
+    default_iterations,
+    default_output_tokens,
+    default_warmup_runs,
+)
+
+
+class LlamaCppBench(Bench):
+    """
+    Benchmark a llama.cpp model
+    """
+
+    unique_name = "llamacpp-bench"
+
+    def __init__(self, monitor_message="Benchmarking LLM"):
+        super().__init__(monitor_message)
+
+        # Don't track memory usage since we are using a llamacpp executable for compute
+        self.save_max_memory_used = False
+
+    @staticmethod
+    def parser(add_help: bool = True) -> argparse.ArgumentParser:
+        parser = __class__.helpful_parser(
+            short_description="Benchmark an LLM in llama.cpp",
+            add_help=add_help,
+        )
+
+        parser = Bench.parser(parser)
+
+        parser.add_argument(
+            "--cli",
+            action="store_true",
+            help="Set this flag to use llama-cli.exe to benchmark model performance. This executable will be called "
+            "once per iteration.  Otherwise, llama-bench.exe is used by default.  In this default behavior behavior, "
+            "the only valid prompt format is integer token lengths. Also, the warmup-iterations parameter is "
+            "ignored and the default value for number of threads is 16.",
+        )
+
+        return parser
+
+    def parse(self, state: State, args, known_only=True) -> argparse.Namespace:
+        """
+        Helper function to parse CLI arguments into the args expected by run()
+        """
+
+        # Call Tool parse method, NOT the Bench parse method
+        parsed_args = Tool.parse(self, state, args, known_only)
+
+        if parsed_args.cli:
+            parsed_args = super().parse(state, args, known_only)
+        else:
+            # Make sure prompts is a list of integers
+            if parsed_args.prompts is None:
+                parsed_args.prompts = [default_prompt_length]
+            prompt_ints = []
+            for prompt_item in parsed_args.prompts:
+                if prompt_item.isdigit():
+                    prompt_ints.append(int(prompt_item))
+                else:
+                    raise Exception(
+                        f"When not using the --cli flag to {self.unique_name}, the prompt format must "
+                        "be in integer format."
+                    )
+            parsed_args.prompts = prompt_ints
+
+        return parsed_args
+
+    def run_prompt(
+        self,
+        state: State,
+        report_progress_fn,
+        prompt: str,
+        iterations: int,
+        warmup_iterations: int,
+        output_tokens: int,
+    ) -> State:
+        """
+        Benchmark llama.cpp model that was loaded by LoadLlamaCpp.
+        """
+
+        if self.first_run_prompt:
+
+            if not hasattr(state, "model") or not isinstance(
+                state.model, LlamaCppAdapter
+            ):
+                raise Exception(
+                    f"{self.__class__.unique_name} requires a LlamaCppAdapter model to be "
+                    "loaded first. Please run load-llama-cpp before this tool."
+                )
+        model: LlamaCppAdapter = state.model
+
+        per_iteration_tokens_per_second = []
+        per_iteration_time_to_first_token = []
+
+        for iteration in range(iterations + warmup_iterations):
+            try:
+                # Use the adapter's generate method which already has the timeout
+                # and error handling
+                model.time_to_first_token = None
+                model.tokens_per_second = None
+                raw_output, stderr = model.generate(
+                    prompt, max_new_tokens=output_tokens, return_raw=True
+                )
+
+                if model.time_to_first_token is None or model.tokens_per_second is None:
+                    error_msg = (
+                        "Could not find timing information in llama.cpp output.\n"
+                    )
+                    error_msg += "Raw output:\n" + raw_output + "\n"
+                    error_msg += "Stderr:\n" + stderr
+                    raise Exception(error_msg)
+
+                self.tokens_out_len_list.append(model.response_tokens)
+
+                if iteration > warmup_iterations - 1:
+                    per_iteration_tokens_per_second.append(model.tokens_per_second)
+                    per_iteration_time_to_first_token.append(model.time_to_first_token)
+
+                report_progress_fn((iteration + 1) / (warmup_iterations + iterations))
+
+            except Exception as e:
+                error_msg = f"Failed to run benchmark: {str(e)}"
+                raise Exception(error_msg)
+
+        self.input_ids_len_list.append(model.prompt_tokens)
+        mean_time_to_first_token = statistics.mean(per_iteration_time_to_first_token)
+        self.mean_time_to_first_token_list.append(mean_time_to_first_token)
+        self.prefill_tokens_per_second_list.append(
+            model.prompt_tokens / mean_time_to_first_token
+        )
+        self.token_generation_tokens_per_second_list.append(
+            statistics.mean(per_iteration_tokens_per_second)
+        )
+        try:
+            self.std_dev_time_to_first_token_list.append(
+                statistics.stdev(per_iteration_time_to_first_token)
+            )
+        except StatisticsError:
+            # Less than 2 measurements
+            self.std_dev_time_to_first_token_list.append(None)
+        try:
+            self.std_dev_token_generation_tokens_per_second_list.append(
+                statistics.stdev(per_iteration_tokens_per_second)
+            )
+        except StatisticsError:
+            # Less than 2 measurements
+            self.std_dev_token_generation_tokens_per_second_list.append(None)
+
+    def run_llama_bench_exe(self, state, prompts, iterations, output_tokens):
+
+        if prompts is None:
+            prompts = [default_prompt_length]
+        elif isinstance(prompts, int):
+            prompts = [prompts]
+
+        state.save_stat("prompts", prompts)
+        state.save_stat("iterations", iterations)
+        state.save_stat("output_tokens", output_tokens)
+
+        model: LlamaCppAdapter = state.model
+        prompt_lengths, pp_tps, pp_tps_sd, tg_tps, tg_tps_sd = model.benchmark(
+            prompts, iterations, output_tokens
+        )
+        self.input_ids_len_list = prompt_lengths
+        self.prefill_tokens_per_second_list = pp_tps
+        if iterations > 1:
+            self.std_dev_prefill_tokens_per_second_list = pp_tps_sd
+        self.mean_time_to_first_token_list = [
+            tokens / tps for tokens, tps in zip(prompt_lengths, pp_tps)
+        ]
+        self.token_generation_tokens_per_second_list = [tg_tps]
+        if iterations > 1:
+            self.std_dev_token_generation_tokens_per_second_list = [tg_tps_sd]
+        self.tokens_out_len_list = [output_tokens] * len(prompts) * iterations
+
+        self.save_stats(state)
+        return state
+
+    def run(
+        self,
+        state: State,
+        prompts: list[str] = None,
+        iterations: int = default_iterations,
+        warmup_iterations: int = default_warmup_runs,
+        output_tokens: int = default_output_tokens,
+        cli: bool = False,
+        **kwargs,
+    ) -> State:
+        """
+        Args:
+            - prompts: List of input prompts used as starting points for LLM text generation
+            - iterations: Number of benchmarking samples to take; results are
+                reported as the median and mean of the samples.
+            - warmup_iterations: Subset of the iterations to treat as warmup,
+                and not included in the results.
+            - output_tokens: Number of new tokens LLM to create.
+            - ggml: Use llama-bench.exe directly
+            - kwargs: Additional parameters used by bench tools
+        """
+
+        # Check that state has the attribute model and it is a LlamaCPP model
+        if not hasattr(state, "model") or not isinstance(state.model, LlamaCppAdapter):
+            raise Exception("Load model using llamacpp-load first.")
+
+        if cli:
+            state = super().run(
+                state, prompts, iterations, warmup_iterations, output_tokens, **kwargs
+            )
+        else:
+            state = self.run_llama_bench_exe(state, prompts, iterations, output_tokens)
+
+        return state
+
+
+# This file was originally licensed under Apache 2.0. It has been modified.
+# Modifications Copyright (c) 2025 AMD

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/src/lemonade/tools/llamacpp/load.py

@@ -93,6 +93,7 @@ class LoadLlamaCpp(FirstTool):
         from lemonade.tools.llamacpp.utils import (
             install_llamacpp,
             get_llama_cli_exe_path,
+            get_llama_bench_exe_path,
             get_llama_installed_version,
             parse_checkpoint,
             download_gguf,
@@ -103,6 +104,8 @@ class LoadLlamaCpp(FirstTool):
 
         install_llamacpp(backend)
 
+        extension = ""
+
         # Check if input is a local folder containing a .GGUF model
         if os.path.isdir(input):
             # input is a local folder
@@ -121,6 +124,17 @@ class LoadLlamaCpp(FirstTool):
                 )
             model_to_use = gguf_files[0]
             full_model_path = os.path.join(local_model_folder, model_to_use)
+            extension = ".gguf"
+
+        elif input.endswith(".gguf") and os.path.isfile(input):
+            # input is a local .gguf file
+            full_model_path = os.path.abspath(input)
+            checkpoint = "local_model"
+            state.checkpoint = checkpoint
+            state.save_stat(Keys.CHECKPOINT, checkpoint)
+            state.save_stat(Keys.LOCAL_MODEL_FOLDER, full_model_path)
+            model_to_use = os.path.basename(full_model_path)
+            extension = ".gguf"
 
         else:
             # Input is a model checkpoint
@@ -161,6 +175,7 @@ class LoadLlamaCpp(FirstTool):
                 model_to_use = os.path.basename(full_model_path)
 
         llama_cli_exe_path = get_llama_cli_exe_path(backend)
+        llama_bench_exe_path = get_llama_bench_exe_path(backend)
         printing.log_info(f"Using llama_cli for GGUF model: {llama_cli_exe_path}")
 
         # Get the directory containing the executable for shared libraries
@@ -174,8 +189,10 @@ class LoadLlamaCpp(FirstTool):
             context_size=context_size,
             threads=threads,
             executable=llama_cli_exe_path,
+            bench_executable=llama_bench_exe_path,
             reasoning=reasoning,
             lib_dir=lib_dir,
+            state=state,
         )
         state.tokenizer = LlamaCppTokenizerAdapter()
         state.device = device
@@ -186,7 +203,9 @@ class LoadLlamaCpp(FirstTool):
             Keys.LLAMA_CLI_VERSION_INFO, get_llama_installed_version(backend)
         )
 
-        status.add_to_state(state=state, name=input, model=model_to_use)
+        status.add_to_state(
+            state=state, name=input, model=model_to_use, extension=extension
+        )
         return state
 
 

{lemonade_sdk-8.1.11 → lemonade_sdk-8.1.12}/src/lemonade/tools/llamacpp/utils.py

@@ -7,6 +7,7 @@ import zipfile
 from typing import Optional
 import subprocess
 import requests
+import lemonade.common.build as build
 import lemonade.common.printing as printing
 from lemonade.tools.adapter import PassthroughTokenizer, ModelAdapter
 
@@ -175,6 +176,13 @@ def get_llama_cli_exe_path(backend: str):
     return get_llama_exe_path("llama-cli", backend)
 
 
+def get_llama_bench_exe_path(backend: str):
+    """
+    Get path to platform-specific llama-bench executable
+    """
+    return get_llama_exe_path("llama-bench", backend)
+
+
 def get_version_txt_path(backend: str):
     """
     Get path to text file that contains version information
@@ -406,6 +414,7 @@ def install_llamacpp(backend):
             exe_paths = [
                 (get_llama_server_exe_path(backend), "llama-server"),
                 (get_llama_cli_exe_path(backend), "llama-cli"),
+                (get_llama_bench_exe_path(backend), "llama-bench"),
             ]
 
             for exe_path, exe_name in exe_paths:
@@ -699,8 +708,10 @@ class LlamaCppAdapter(ModelAdapter):
         context_size,
         threads,
         executable,
+        bench_executable,
         reasoning=False,
         lib_dir=None,
+        state=None,
     ):
         super().__init__()
 
@@ -712,8 +723,10 @@ class LlamaCppAdapter(ModelAdapter):
         self.context_size = context_size
         self.threads = threads
         self.executable = os.path.normpath(executable)
+        self.bench_executable = os.path.normpath(bench_executable)
         self.reasoning = reasoning
         self.lib_dir = lib_dir
+        self.state = state
 
     def generate(
         self,
@@ -754,32 +767,54 @@ class LlamaCppAdapter(ModelAdapter):
             self.executable,
             "-m",
             self.model,
-            "--ctx-size",
+            "--ctx-size",  # size of the prompt context, 0 = loaded from model
             str(self.context_size),
-            "-n",
+            "-n",  # number of tokens to predict, -1 = infinity, =2 - until context filled
             str(n_predict),
-            "-t",
+            "-t",  # number of threads to use during generation
             str(self.threads),
             "-p",
             prompt,
+            "-b",  # logical maximum batch size
+            "1",
+            "-ub",  # physical maximum batch size
+            "1",
             "--temp",
             str(temperature),
             "--top-p",
             str(top_p),
             "--top-k",
             str(top_k),
-            "-e",
-            "-no-cnv",
-            "--reasoning-format",
+            "-e",  # process escape sequences
+            "--no-conversation",  # disable conversation mode
+            "--reasoning-format",  # leaves thoughts unparsed in message content
             "none",
         ]
 
+        # If prompt exceeds 500 characters, then use a file
+        if len(prompt) < 500:
+            cmd += ["-p", prompt]
+        else:
+            # Create prompt file in cache directory
+            prompt_file = os.path.join(
+                build.output_dir(self.state.cache_dir, self.state.build_name),
+                "prompt.txt",
+            )
+            with open(prompt_file, "w", encoding="utf-8") as file:
+                file.write(prompt)
+            cmd += ["-f", prompt_file]
+
         # Configure GPU layers: 99 for GPU, 0 for CPU-only
         ngl_value = "99" if self.device == "igpu" else "0"
         cmd = cmd + ["-ngl", ngl_value]
 
         cmd = [str(m) for m in cmd]
 
+        # save llama-cli command
+        self.state.llama_cli_cmd = getattr(self.state, "llama_cli_cmd", []) + [
+            " ".join(cmd)
+        ]
+
         try:
             # Set up environment with library path for Linux
             env = os.environ.copy()
@@ -809,6 +844,15 @@ class LlamaCppAdapter(ModelAdapter):
             )
 
             raw_output, stderr = process.communicate(timeout=600)
+
+            # save llama-cli command output with performance info to state
+            # (can be viewed in state.yaml file in cache)
+            self.state.llama_cli_stderr = getattr(
+                self.state, "llama_cli_stderr", []
+            ) + [
+                [line for line in stderr.splitlines() if line.startswith("llama_perf_")]
+            ]
+
             if process.returncode != 0:
                 error_msg = f"llama.cpp failed with return code {process.returncode}.\n"
                 error_msg += f"Command: {' '.join(cmd)}\n"
@@ -873,7 +917,108 @@ class LlamaCppAdapter(ModelAdapter):
             return [output_text]
 
         except Exception as e:
-            error_msg = f"Failed to run llama.cpp command: {str(e)}\n"
+            error_msg = f"Failed to run llama-cli.exe command: {str(e)}\n"
+            error_msg += f"Command: {' '.join(cmd)}"
+            raise Exception(error_msg)
+
+    def benchmark(self, prompts, iterations, output_tokens):
+        """
+        Runs the llama-bench.exe tool to measure TTFT and TPS
+        """
+        cmd = [
+            self.bench_executable,
+            "-m",
+            self.model,
+            "-r",
+            iterations,
+            "-p",
+            ",".join([str(p) for p in prompts]),
+            "-n",
+            output_tokens,
+            "-t",
+            self.threads if self.threads > 0 else 16,
+            "-b",
+            1,
+            "-ub",
+            1,
+        ]
+        cmd = [str(m) for m in cmd]
+
+        # save llama-bench command
+        self.state.llama_bench_cmd = " ".join(cmd)
+
+        try:
+            # Set up environment with library path for Linux
+            env = os.environ.copy()
+
+            # Load environment variables from .env file in the executable directory
+            exe_dir = os.path.dirname(self.executable)
+            env_file_path = os.path.join(exe_dir, ".env")
+            if os.path.exists(env_file_path):
+                load_dotenv(env_file_path, override=True)
+                env.update(os.environ)
+
+            if self.lib_dir and os.name != "nt":  # Not Windows
+                current_ld_path = env.get("LD_LIBRARY_PATH", "")
+                if current_ld_path:
+                    env["LD_LIBRARY_PATH"] = f"{self.lib_dir}:{current_ld_path}"
+                else:
+                    env["LD_LIBRARY_PATH"] = self.lib_dir
+
+            process = subprocess.Popen(
+                cmd,
+                stdout=subprocess.PIPE,
+                stderr=subprocess.PIPE,
+                universal_newlines=True,
+                encoding="utf-8",
+                errors="replace",
+                env=env,
+            )
+
+            raw_output, stderr = process.communicate(timeout=600)
+
+            # save llama-bench command output with performance info to state
+            # (can be viewed in state.yaml file in cache)
+            self.state.llama_bench_standard_output = raw_output.splitlines()
+
+            if process.returncode != 0:
+                error_msg = (
+                    f"llama-bench.exe failed with return code {process.returncode}.\n"
+                )
+                error_msg += f"Command: {' '.join(cmd)}\n"
+                error_msg += f"Error output:\n{stderr}\n"
+                error_msg += f"Standard output:\n{raw_output}"
+                raise Exception(error_msg)
+
+            if raw_output is None:
+                raise Exception("No output received from llama-bench.exe process")
+
+            # Parse information from llama-bench.exe output
+            prompt_lengths = []
+            pp_tps = []
+            pp_tps_sd = []
+            tg_tps = None
+            tg_tps_sd = None
+
+            for line in self.state.llama_bench_standard_output:
+                # Parse TPS information
+                for p in prompts:
+                    if f"pp{p:d}" in line:
+                        parts = line.split("|")
+                        timings = parts[-2].strip().split(" ")
+                        prompt_lengths.append(p)
+                        pp_tps.append(float(timings[0]))
+                        pp_tps_sd.append(float(timings[-1]))
+                    if f"tg{output_tokens:d}" in line:
+                        parts = line.split("|")
+                        timings = parts[-2].strip().split(" ")
+                        tg_tps = float(timings[0])
+                        tg_tps_sd = float(timings[-1])
+
+            return prompt_lengths, pp_tps, pp_tps_sd, tg_tps, tg_tps_sd
+
+        except Exception as e:
+            error_msg = f"Failed to run llama-bench.exe command: {str(e)}\n"
             error_msg += f"Command: {' '.join(cmd)}"
             raise Exception(error_msg)