Merge branch 'main' into more_flexible_tdt_durations

nemo/deploy/nlp/__init__.py

@@ -15,7 +15,7 @@
 
 use_query_llm = True
 try:
-    from nemo.deploy.nlp.query_llm import NemoQueryLLM
+    from nemo.deploy.nlp.query_llm import NemoQueryLLM, NemoQueryLLMPyTorch
 except Exception:
     use_query_llm = False
 

nemo/deploy/nlp/megatronllm_deployable.py

@@ -15,6 +15,7 @@
 import logging
 from enum import IntEnum, auto
 from pathlib import Path
+from typing import List
 
 import numpy as np
 import torch
@@ -129,6 +130,12 @@ def _load_from_nemo_checkpoint(self, nemo_checkpoint_filepath: str, num_devices:
                 nemo_checkpoint_filepath, trainer=trainer, return_config=True
             )
             # transformer_engine should always be true according to EricH, but GPT-2B model will fail if it is enabled
+            if not custom_config.transformer_engine:
+                LOGGER.warning(
+                    "MegatronLLMDeployable expects model config transformer_engine=True, but this model has it =False. "
+                    "Overriding it to =True, but this may break certain checkpoints converted on older Nemo versions. "
+                    "If your model breaks, please try re-converting the checkpoint on the current Nemo version."
+                )
             custom_config.transformer_engine = True
             # using multi-gpu for tensor parallelism directly for now, could do pipeline parallel instead or a combination
             custom_config.tensor_model_parallel_size = num_devices
@@ -233,9 +240,7 @@ def _length_params_from_triton_inputs(**inputs: np.ndarray):
                 length_params[length_param_field] = inputs.pop(length_param_field)[0][0]
         return length_params
 
-    @batch
-    def triton_infer_fn(self, **inputs: np.ndarray):
-        """Triton server inference function that actually runs the model"""
+    def generate(self, inputs: List[str], length_params: LengthParam, sampling_params: SamplingParam):
         if torch.distributed.is_initialized():
             distributed_rank = torch.distributed.get_rank()
             if distributed_rank != 0:
@@ -245,13 +250,16 @@ def triton_infer_fn(self, **inputs: np.ndarray):
             signal_value = ServerSync.SIGNAL.to_long_tensor()
             torch.distributed.broadcast(signal_value, 0)
 
+        return self.model.generate(inputs=inputs, length_params=length_params, sampling_params=sampling_params)
+
+    @batch
+    def triton_infer_fn(self, **inputs: np.ndarray):
+        """Triton server inference function that actually runs the model"""
         input_strings = str_ndarray2list(inputs.pop("prompts"))
         sampling_params = self._sampling_params_from_triton_inputs(**inputs)
         length_params = self._length_params_from_triton_inputs(**inputs)
 
-        model_output = self.model.generate(
-            inputs=input_strings, length_params=length_params, sampling_params=sampling_params
-        )
+        model_output = self.generate(input_strings, length_params, sampling_params)
         '''
             model_output['sentences'] will be a list of strings (one per prompt)
             other fields will either be a list of lists (tokens, for example)

nemo/deploy/nlp/query_llm.py

@@ -30,23 +30,99 @@ def __init__(self, url, model_name):
         self.url = url
         self.model_name = model_name
 
-    @abstractmethod
+
+class NemoQueryLLMPyTorch(NemoQueryLLMBase):
+    """
+    Sends a query to Triton for LLM inference
+
+    Example:
+        from nemo.deploy import NemoTritonQueryLLMPyTorch
+
+        nq = NemoTritonQueryLLMPyTorch(url="localhost", model_name="GPT-2B")
+
+        prompts = ["hello, testing GPT inference", "another GPT inference test?"]
+        output = nq.query_llm(
+            prompts=prompts,
+            max_length=100,
+            top_k=1,
+            top_p=0.0,
+            temperature=0.0,
+        )
+        print("prompts: ", prompts)
+    """
+
+    def __init__(self, url, model_name):
+        super().__init__(
+            url=url,
+            model_name=model_name,
+        )
+
+    # these arguments are explicitly defined in order to make it clear to user what they can pass
+    # names and optionality should exactly match the get_triton_input() results for MegatronGPTDeployable
     def query_llm(
         self,
         prompts,
-        stop_words_list=None,
-        bad_words_list=None,
-        no_repeat_ngram_size=None,
-        max_output_len=512,
-        top_k=1,
-        top_p=0.0,
-        temperature=1.0,
-        random_seed=None,
-        task_id=None,
-        lora_uids=None,
+        use_greedy: bool = None,
+        temperature: float = None,
+        top_k: int = None,
+        top_p: float = None,
+        repetition_penalty: float = None,
+        add_BOS: bool = None,
+        all_probs: bool = None,
+        compute_logprob: bool = None,
+        end_strings=None,
+        min_length: int = None,
+        max_length: int = None,
         init_timeout=60.0,
     ):
-        pass
+        """
+        Query the Triton server synchronously and return a list of responses.
+
+        Args:
+            prompts (List(str)): list of sentences.
+            use_greedy (bool): use greedy sampling, effectively the same as top_k=1
+            temperature (float): A parameter of the softmax function, which is the last layer in the network.
+            top_k (int): limits us to a certain number (K) of the top tokens to consider.
+            top_p (float): limits us to the top tokens within a certain probability mass (p).
+            repetition_penalty (float): penalty applied to repeated sequences, 1.0 means no penalty.
+            add_BOS (bool): whether or not to add a BOS (beginning of sentence) token.
+            all_probs (bool): when using compute_logprob, returns probabilities for all tokens in vocabulary.
+            compute_logprob (bool): get back probabilities of all tokens in the sequence.
+            end_strings (List(str)): list of strings which will terminate generation when they appear in the output.
+            min_length (int): min generated tokens.
+            max_length (int): max generated tokens.
+            init_timeout (flat): timeout for the connection.
+        """
+        prompts = str_list2numpy(prompts)
+        inputs = {
+            "prompts": prompts,
+        }
+        if use_greedy is not None:
+            inputs["use_greedy"] = np.full(prompts.shape, use_greedy, dtype=np.bool_)
+        if temperature is not None:
+            inputs["temperature"] = np.full(prompts.shape, temperature, dtype=np.single)
+        if top_k is not None:
+            inputs["top_k"] = np.full(prompts.shape, top_k, dtype=np.int_)
+        if top_p is not None:
+            inputs["top_p"] = np.full(prompts.shape, top_p, dtype=np.single)
+        if repetition_penalty is not None:
+            inputs["repetition_penalty"] = np.full(prompts.shape, repetition_penalty, dtype=np.single)
+        if add_BOS is not None:
+            inputs["add_BOS"] = np.full(prompts.shape, add_BOS, dtype=np.bool_)
+        if all_probs is not None:
+            inputs["all_probs"] = np.full(prompts.shape, all_probs, dtype=np.bool_)
+        if compute_logprob is not None:
+            inputs["compute_logprob"] = np.full(prompts.shape, compute_logprob, dtype=np.bool_)
+        if end_strings is not None:
+            inputs["end_strings"] = str_list2numpy(end_strings)
+        if min_length is not None:
+            inputs["min_length"] = np.full(prompts.shape, min_length, dtype=np.int_)
+        if max_length is not None:
+            inputs["max_length"] = np.full(prompts.shape, max_length, dtype=np.int_)
+
+        with ModelClient(self.url, self.model_name, init_timeout_s=init_timeout) as client:
+            result_dict = client.infer_batch(**inputs)
+            return result_dict
 
 
 class NemoQueryLLM(NemoQueryLLMBase):

scripts/deploy/nlp/deploy_inframework_triton.py

@@ -0,0 +1,103 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import logging
+import sys
+
+from nemo.deploy import DeployPyTriton
+
+LOGGER = logging.getLogger("NeMo")
+
+megatron_llm_supported = True
+try:
+    from nemo.deploy.nlp import MegatronLLMDeployable
+except Exception as e:
+    LOGGER.warning(f"Cannot import MegatronLLMDeployable, it will not be available. {type(e).__name__}: {e}")
+    megatron_llm_supported = False
+
+
+def get_args(argv):
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+        description=f"Deploy nemo models to Triton",
+    )
+    parser.add_argument("-nc", "--nemo_checkpoint", type=str, help="Source .nemo file")
+    parser.add_argument("-tmn", "--triton_model_name", required=True, type=str, help="Name for the service")
+    parser.add_argument("-tmv", "--triton_model_version", default=1, type=int, help="Version for the service")
+    parser.add_argument(
+        "-trp", "--triton_port", default=8000, type=int, help="Port for the Triton server to listen for requests"
+    )
+    parser.add_argument(
+        "-tha", "--triton_http_address", default="0.0.0.0", type=str, help="HTTP address for the Triton server"
+    )
+    parser.add_argument("-ng", "--num_gpus", default=1, type=int, help="Number of GPUs for the deployment")
+    parser.add_argument("-mbs", "--max_batch_size", default=8, type=int, help="Max batch size of the model")
+    parser.add_argument("-dm", "--debug_mode", default=False, action='store_true', help="Enable debug mode")
+    args = parser.parse_args(argv)
+    return args
+
+
+def get_nemo_deployable(args):
+    if args.nemo_checkpoint is None:
+        raise ValueError("In-Framework deployment requires a .nemo checkpoint")
+
+    return MegatronLLMDeployable(args.nemo_checkpoint, args.num_gpus)
+
+
+def nemo_deploy(argv):
+    args = get_args(argv)
+
+    if args.debug_mode:
+        loglevel = logging.DEBUG
+    else:
+        loglevel = logging.INFO
+
+    LOGGER.setLevel(loglevel)
+    LOGGER.info("Logging level set to {}".format(loglevel))
+    LOGGER.info(args)
+
+    if not megatron_llm_supported:
+        raise ValueError("MegatronLLMDeployable is not supported in this environment.")
+    triton_deployable = get_nemo_deployable(args)
+
+    try:
+        nm = DeployPyTriton(
+            model=triton_deployable,
+            triton_model_name=args.triton_model_name,
+            triton_model_version=args.triton_model_version,
+            max_batch_size=args.max_batch_size,
+            port=args.triton_port,
+            address=args.triton_http_address,
+        )
+
+        LOGGER.info("Triton deploy function will be called.")
+        nm.deploy()
+    except Exception as error:
+        LOGGER.error("Error message has occurred during deploy function. Error message: " + str(error))
+        return
+
+    try:
+        LOGGER.info("Model serving on Triton is will be started.")
+        nm.serve()
+    except Exception as error:
+        LOGGER.error("Error message has occurred during deploy function. Error message: " + str(error))
+        return
+
+    LOGGER.info("Model serving will be stopped.")
+    nm.stop()
+
+
+if __name__ == '__main__':
+    nemo_deploy(sys.argv[1:])

scripts/deploy/nlp/query_inframework.py

@@ -0,0 +1,83 @@
+# Copyright (c) 2023, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import sys
+
+from nemo.deploy.nlp.query_llm import NemoQueryLLMPyTorch
+
+
+def get_args(argv):
+    parser = argparse.ArgumentParser(
+        formatter_class=argparse.ArgumentDefaultsHelpFormatter,
+        description=f"Queries Triton server running an in-framework Nemo model",
+    )
+    parser.add_argument("-u", "--url", default="0.0.0.0", type=str, help="url for the triton server")
+    parser.add_argument("-mn", "--model_name", required=True, type=str, help="Name of the triton model")
+    prompt_group = parser.add_mutually_exclusive_group(required=True)
+    prompt_group.add_argument("-p", "--prompt", required=False, type=str, help="Prompt")
+    prompt_group.add_argument("-pf", "--prompt_file", required=False, type=str, help="File to read the prompt from")
+    parser.add_argument("-mol", "--max_output_len", default=128, type=int, help="Max output token length")
+    parser.add_argument("-tk", "--top_k", default=1, type=int, help="top_k")
+    parser.add_argument("-tpp", "--top_p", default=0.0, type=float, help="top_p")
+    parser.add_argument("-t", "--temperature", default=1.0, type=float, help="temperature")
+    parser.add_argument("-it", "--init_timeout", default=60.0, type=float, help="init timeout for the triton server")
+
+    args = parser.parse_args(argv)
+    return args
+
+
+def query_llm(
+    url,
+    model_name,
+    prompts,
+    max_output_len=128,
+    top_k=1,
+    top_p=0.0,
+    temperature=1.0,
+    init_timeout=60.0,
+):
+    nemo_query = NemoQueryLLMPyTorch(url, model_name)
+    return nemo_query.query_llm(
+        prompts=prompts,
+        max_length=max_output_len,
+        top_k=top_k,
+        top_p=top_p,
+        temperature=temperature,
+        init_timeout=init_timeout,
+    )
+
+
+def query(argv):
+    args = get_args(argv)
+
+    if args.prompt_file is not None:
+        with open(args.prompt_file, "r") as f:
+            args.prompt = f.read()
+
+    outputs = query_llm(
+        url=args.url,
+        model_name=args.model_name,
+        prompts=[args.prompt],
+        max_output_len=args.max_output_len,
+        top_k=args.top_k,
+        top_p=args.top_p,
+        temperature=args.temperature,
+        init_timeout=args.init_timeout,
+    )
+    print(outputs["sentences"][0][0])
+
+
+if __name__ == '__main__':
+    query(sys.argv[1:])

tests/deploy/nemo_deploy.py

@@ -27,7 +27,7 @@
 run_export_tests = True
 try:
     from nemo.deploy import DeployPyTriton
-    from nemo.deploy.nlp import NemoQueryLLM
+    from nemo.deploy.nlp import NemoQueryLLM, NemoQueryLLMPyTorch
     from nemo.export import TensorRTLLM
 except Exception as e:
     run_export_tests = False
@@ -140,7 +140,7 @@ def run_in_framework_inference(
     )
     nm.deploy()
     nm.run()
-    nq = NemoQueryLLM(url="localhost:8000", model_name=model_name)
+    nq = NemoQueryLLMPyTorch(url="localhost:8000", model_name=model_name)
 
     output_deployed = nq.query_llm(
         prompts=prompt,