Generate: end-to-end compilation

[gante]

What does this PR do?

Closes #29374
Closes #29151

This PR introduces an MVP for end-to-end generate compilation -- the whole generate function can be compiled into a single graph. From this state, we can start testing and improving many aspects of our generate code wrt compilation, as well as unlock features that rely on generate being a single graph.

⚠️ generate compilation in this PR has MANY restrictions, including:

• Compilation is veeery slow 🐌 ;
• torch.multinomial(probs, num_samples=1) can't be compiled into a cuda graph, so do_sample=True is not compatible without a cuda graph break;
• stopping when the model throws an EOS token is not compatible;

---

Tests

The following slow tests were passing (as of 4241ab3):

• RUN_SLOW=1 py.test -vv tests/models/ -k test_generate_compile_fullgraph
• RUN_SLOW=1 py.test -vv tests/utils/test_cache_utils.py
• RUN_SLOW=1 py.test -vv tests/models/llama/test_modeling_llama.py

---

Performance

TL;DR:

• no big speed wins vs compiling forward, ~5% faster. The main benefit is the ability to export a single graph
• compilation takes a LOT of time. On both test devices, the process got killed when max_new_tokens=256

Setup:

• A100 80GB / RTX 3090 24GB
• compiling forward but calling generate (i.e. there is some overhead from calling the uncompiled generate)
• model: google/gemma-2b

code

import copy
import os
import torch
from torch.utils import benchmark

from transformers import AutoTokenizer, AutoModelForCausalLM, StaticCache

os.environ["TOKENIZERS_PARALLELISM"] = "false"

# Benchmarking settings
BSZ = [1, 4]
NEW_TOK = [16, 256]
# BSZ = [1]
# NEW_TOK = [16]
N_ITER = 20
MODEL_ID = "google/gemma-2b"
ATTN_IMPL = "sdpa"

# Other constants
FRANCE_ARTICLE = (  # @noqa
    """<s>Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. \"One can hear cries of 'My God' in several languages,\" Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object.  Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France's accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren't revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he's accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz's battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz's correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz's possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and "held all the licenses required." Earlier, a spokesman for the prosecutor\'s office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot's license. Kumpa emphasized there's no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot's license, a European government official briefed on the investigation told CNN on Tuesday. While flying was "a big part of his life," the source said, it\'s only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz's girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there's more to the story, said Brian Russell, a forensic psychologist. "Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren't going to keep doing their job and they're upset about that and so they're suicidal," he said. "But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person's problems." Germanwings crash compensation: What we know . Who was the captain of Germanwings Flight 9525? CNN's Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN's Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel and Anna-Maja Rappard contributed to this report."""
)


tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, padding_side="left")
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, torch_dtype=torch.float16, attn_implementation=ATTN_IMPL
).to("cuda")
prompt_length = tokenizer([FRANCE_ARTICLE], return_tensors="pt").input_ids.shape[1]
label_ms_per_token = f"Throughput (time/foward pass, prompt = {prompt_length} tokens)"
label_first_step = f"First call (time, prompt = {prompt_length} tokens)"


def print_results(all_results):
    print("\n")
    compare = benchmark.Compare(all_results)
    compare.trim_significant_figures()
    compare.colorize(rowwise = True)
    compare.print()


def time_generate_call(model, task, ms_per_token, first_step, static_cache=False):
    for bsz in BSZ:
        for max_new_tokens in NEW_TOK:
            input_ids = tokenizer([FRANCE_ARTICLE] * bsz, return_tensors="pt").to("cuda")
            description = f"batch size, max_new_tokens: {bsz, max_new_tokens}"
            task_spec_ms_per_token = benchmark.TaskSpec(
                stmt="", setup="", description=task, label=label_ms_per_token, sub_label=description
            )
            task_spec_ms_first_step = benchmark.TaskSpec(
                stmt="", setup="", description=task, label=label_first_step, sub_label=description
            )

            # generate EXACTLY `max_new_tokens` tokens (no early termination due to `eos_token_id`)
            generation_kwargs = {
                "max_new_tokens": max_new_tokens,
                "min_new_tokens": max_new_tokens,
                "eos_token_id": None,
                "do_sample": False,
            }
            if static_cache:
                generation_kwargs["cache_implementation"] = "static"
            generation_config = copy.deepcopy(model.generation_config)
            generation_config.update(**generation_kwargs)

            past_key_values = None

            torch.compiler.reset()
            results = []
            for _ in range(N_ITER):
                start = torch.cuda.Event(enable_timing=True)
                end = torch.cuda.Event(enable_timing=True)
                start.record()
                gen_out = model.generate(
                    **input_ids, generation_config=generation_config, past_key_values=past_key_values
                )
                end.record()
                torch.cuda.synchronize()
                total_time = start.elapsed_time(end) / 1000  # time in seconds
                time_per_forward = total_time / max_new_tokens
                assert gen_out.shape[1] == max_new_tokens + prompt_length
                results.append(time_per_forward)

            if static_cache:
                del model._cache

            ms_per_token.append(benchmark.Measurement(1, results[3:], task_spec_ms_per_token, metadata=None))
            first_step.append(benchmark.Measurement(
                1, [results[0] * max_new_tokens], task_spec_ms_first_step, metadata=None)
            )
            print_results(ms_per_token)
            print_results(first_step)
            print("*" * 80)


ms_per_token = []
first_step = []

# dynamic
time_generate_call(model, "dynamic", ms_per_token, first_step)

# static
time_generate_call(model, "static", ms_per_token, first_step, static_cache=True)

# static + forward compiled
torch.compiler.reset()
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)
time_generate_call(model, "static + fwd compiled", ms_per_token, first_step, static_cache=True)

# generate compiled
torch.compiler.reset()
model.generate = torch.compile(model.generate, mode="reduce-overhead", fullgraph=True)
time_generate_call(model, "generate compiled", ms_per_token, first_step)

A100

RTX3090

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ydshieh]

This is great 🔥 🚀 . Please ping me once ready - I would like to run some benchmarks!

Also, it would be best to have an integration test for at least 2 models, probably llama and gemma.

[ydshieh]

I added [WIP] to the title but didn't change it to draft mode. Hope I am doing it right.

[ydshieh]

@gante Do you have the generation running time numbers of forward compile v.s. end-to-end compile, say for llama?

[gante]

@ydshieh I'm not sure it will be faster, the biggest win is in terms of compatibility (compiled graphs are more portable and easily consumed by specialized hardware). I will run and share a few benchmarks in any case 🤗

[ydshieh]

Nice! If you are OK with that, maybe adopt the following small & simple script for a (even just a first) benchmark.

import os
import torch
import datetime

from transformers import AutoTokenizer, AutoModelForCausalLM

token = "ADD_YOUR_OWN_TOKEN"

os.environ["TOKENIZERS_PARALLELISM"] = "false"

batch_size = 1
n_iter = 3

ckpt = "google/gemma-2b"

tokenizer = AutoTokenizer.from_pretrained(ckpt, token=token)
model = AutoModelForCausalLM.from_pretrained(ckpt, token=token, torch_dtype=torch.float16).to("cuda")

model.generation_config.max_new_tokens = 16
model.generation_config.max_new_tokens = 16

model.generation_config.cache_implementation = "static"
model.forward = torch.compile(model.forward, mode="reduce-overhead", fullgraph=True)

input_text = "Why dogs are cute."
input_ids = tokenizer([input_text] * batch_size, return_tensors="pt").to("cuda")

for i in range(n_iter):
    s = datetime.datetime.now()
    outputs = model.generate(**input_ids, do_sample=False)
    t = datetime.datetime.now()
    e = (t-s).total_seconds()
    print(e)

[gante]

@ydshieh TL;DR not much faster, at least for now :) (benchmarks in the PR header)

[zucchini-nlp]

Very cool 🔥

[gante]

(seemingly unrelated CI failures)

[ArthurZucker]

I'll have a look!

[ArthurZucker]

🔥 will this be breaking compatibility for people who use torch.compile(model.forward)

[ArthurZucker]

kinda wondering if torch compile fairs well with ints like this, but I gues yeah

[ArthurZucker]

maybe we can open an issue on torch directly!

[ydshieh]

compilation takes a LOT of time. On both test devices, the process got killed when max_new_tokens=256

It might relate to pytorch/pytorch#128424. But max_new_tokens=256 is quite short, however in this PR, there are much more involved (the code from generate but outside forward).

The process got killed: this is likely due to memory? I also observed this and reported in the above pytorch issue page.

Do you try compile without reduce-overhead? I know it's not ideal, but just to see if it works much better (except the ultra performance)

[gante]

(ready to be merged, needs CI fixing from external issues -- on it)

[ArthurZucker]

Very good, just don't like the control flow on the modeling size as we should update the modeling code IMO and the prepare_inputs_for generation should forget about get seq len and get max lenght if possible (infered from generation config? )

[gante]

Ran benchmarks again to double-check, the conclusions at the top persist:

(gemma 2b, measured on RTX4090)