KoboldAI-Client/tpu_mtj_backend.py

1420 lines
67 KiB
Python

'''
This file is AGPL-licensed.
Some of the code in this file is from Clover Edition:
https://github.com/cloveranon/Clover-Edition/blob/master/aidungeon/gpt2generator.py
The license for Clover Edition is shown below:
Copyright (c) 2019 Nick Walton
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
'''
import utils
import multiprocessing
from typing import Any, Callable, Dict, List, NamedTuple, Optional, Tuple, TypeVar
import progressbar
import time
import os
import sys
import json
import zipfile
import requests
import random
import jax
import jax.dlpack
from jax.config import config
from jax.experimental import maps
import jax.numpy as jnp
import numpy as np
import haiku as hk
from transformers import AutoTokenizer, GPT2Tokenizer, AutoModelForCausalLM, GPTNeoForCausalLM
from tokenizers import Tokenizer
from mesh_transformer.checkpoint import read_ckpt_lowmem
from mesh_transformer.transformer_shard import CausalTransformer, CausalTransformerShard, PlaceholderTensor
from mesh_transformer.util import to_bf16
params: Dict[str, Any] = {}
__seed = random.randrange(2**64)
rng = random.Random(__seed)
def get_rng_seed():
return __seed
def set_rng_seed(seed: int):
global __seed, rng
rng = random.Random(seed)
__seed = seed
return seed
def randomize_rng_seed():
return set_rng_seed(random.randrange(2**64))
def get_rng_state():
return rng
def set_rng_state(state):
global rng
rng = state
def new_rng_state(seed: int):
return random.Random(seed)
def warper_callback(logits) -> np.array:
raise NotImplementedError("`tpu_mtj_backend.warper_callback()` needs to be defined")
def stopping_callback(generated, n_generated, excluded_world_info) -> Tuple[List[set], bool, bool]:
raise NotImplementedError("`tpu_mtj_backend.stopping_callback()` needs to be defined")
def settings_callback() -> dict:
return {
"sampler_order": utils.default_sampler_order.copy(),
"top_p": 0.9,
"temp": 0.5,
"top_k": 0,
"tfs": 1.0,
"typical": 1.0,
"top_a": 0.0,
"repetition_penalty": 1.0,
"rpslope": 0.0,
"rprange": 0,
}
def started_compiling_callback() -> None:
pass
def stopped_compiling_callback() -> None:
pass
def compiling_callback() -> None:
pass
def show_spinner():
bar = progressbar.ProgressBar(max_value=progressbar.UnknownLength, widgets=[progressbar.Timer(), ' ', progressbar.BouncingBar(left='[', right=']', marker='')])
i = 0
while True:
bar.update(i)
time.sleep(0.1)
i += 1
__F = TypeVar("__F", bound=Callable)
__T = TypeVar("__T")
def __move_xmap(f: __F, out_axis: str) -> __F:
return maps.xmap(
f,
in_axes=(["shard", ...], ["batch", ...]),
out_axes=[out_axis, ...],
axis_resources={'shard': 'mp', 'batch': 'dp'},
)
def __shard_xmap(batch_dim=1):
xmap = __move_xmap(lambda s, b: s, "shard")
def inner(x: __T) -> __T:
return xmap(x, np.empty(batch_dim))
return inner
def __batch_xmap(shard_dim=1):
xmap = __move_xmap(lambda s, b: b, "batch")
def inner(x: __T) -> __T:
return xmap(np.empty(shard_dim), x)
return inner
class _EmptyState(NamedTuple):
pass
class _DummyOptimizer:
def init(*args, **kwargs):
return _EmptyState()
def apply_repetition_penalty_dynamic(logits, tokens, repetition_penalty, generated_index, gen_length, rpslope, rprange):
'''
This gets called by generate_loop_fn to apply repetition penalty
to the 1D array logits using the provided 1D array of tokens to penalize
'''
tokens = np.minimum(tokens, params["n_vocab"]-1) # https://github.com/google/jax/issues/3774
rpslope = np.int32(rpslope)
rprange = np.int32(rprange)
clipped_rprange = rprange if rprange > 0 else tokens.shape[-1]
penalty_arange = np.roll(np.arange(tokens.shape[-1]) + (clipped_rprange - tokens.shape[-1]), generated_index, axis=-1)
# Make a new array with the same length as the tokens array but with
# each element replaced by the value at the corresponding index in the
# logits array; e.g.
# if logits is [77, 5, 3, 98] and tokens is [0, 1, 2, 3, 2, 3, 1],
# then penalty_logits will be [77, 5, 3, 98, 3, 98, 5]
penalty_logits = np.take(logits, tokens)
# Repetition penalty slope
if rpslope != 0.0 and rprange > 0:
_penalty = (penalty_arange/(rprange - 1)) * 2 - 1
_penalty = (rpslope * _penalty) / (1 + np.abs(_penalty) * (rpslope - 1))
_penalty = 1 + ((_penalty + 1) / 2) * (repetition_penalty - 1)
repetition_penalty = _penalty
# Divide positive values by repetition_penalty and multiply negative
# values by repetition_penalty (the academic publication that described
# this technique actually just only divided, but that would cause tokens
# with negative logits to become more likely, which is obviously wrong)
penalty_logits = np.where(
penalty_arange >= 0,
np.where(
penalty_logits > 0,
penalty_logits/repetition_penalty,
penalty_logits*repetition_penalty,
),
penalty_logits,
)
# Finally, put those penalized logit values back into their original
# positions in the logits array
logits[tokens] = penalty_logits
return logits
def kobold_sample_dynamic(key, logits, rpargs, sampler_order: Optional[np.ndarray] = None, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0, top_a=0.0):
'''
This gets called by generate_loop_fn to apply a series of 6 filters
to the logits (top-k, then top-a, then top-p, then TFS, then typical, then temperature)
before picking one token using the modified logits
'''
# Top-k (keep only the k tokens with the highest logits and remove
# the rest, by setting their logits to negative infinity)
def top_k_filter(logits):
# After sorting the logits array in descending order,
# sorted_indices_to_remove is a 1D array that is True for tokens
# in the sorted logits array we want to remove and False for ones
# we want to keep, in this case the first top_k elements will be
# False and the rest will be True
sorted_indices_to_remove = np.arange(len(logits)) >= top_k
# Unsort the logits array back to its original configuration and
# remove tokens we need to remove
_, indices_to_remove = jax.lax.sort_key_val(
np.argsort(-logits),
sorted_indices_to_remove,
)
return np.where(indices_to_remove, -np.inf, logits)
# Top-a (remove all tokens that have softmax probability less than
# a*m^2 where m is the maximum softmax probability)
def top_a_filter(logits):
# Replace every element in the logits array
# with e (Euler's number) to the power of that element, and divide
# each element of the new array by the sum of the elements in the
# new array
probabilities = np.array(jax.nn.softmax(logits), copy=True)
# Find the largest probability
probs_max = probabilities.max()
# Remove tokens
return np.where(probabilities < probs_max * probs_max * top_a, -np.inf, logits)
# Top-p (after sorting the remaining tokens again in descending order of
# logit, remove the ones that have cumulative softmax probability
# greater than p)
def top_p_filter(logits):
# Sort the logits array in descending order, replace every element
# with e (Euler's number) to the power of that element, and divide
# each element of the new array by the sum of the elements in the
# new array
sorted_logits = -np.sort(-logits)
probabilities = np.array(jax.nn.softmax(sorted_logits), copy=True)
# Calculate cumulative_probabilities as the prefix-sum array of
# probabilities
cumulative_probabilities = np.cumsum(probabilities, axis=-1)
# We want to remove tokens with cumulative probability higher
# than top_p
sorted_indices_to_remove = cumulative_probabilities > top_p
# Don't ever remove the token with the highest logit, even if
# the probability is higher than top_p
sorted_indices_to_remove[0] = False
# Unsort and remove
_, indices_to_remove = jax.lax.sort_key_val(
np.argsort(-logits),
sorted_indices_to_remove,
)
return np.where(indices_to_remove, -np.inf, logits)
# Tail free sampling (basically top-p a second time on remaining tokens
# except it's the "cumulative normalized absolute second finite
# differences of the softmax probabilities" instead of just the
# cumulative softmax probabilities)
def tail_free_filter(logits):
# Sort in descending order
sorted_logits = -np.sort(-logits)
# Softmax again
probabilities = np.array(jax.nn.softmax(sorted_logits), copy=True)
# Calculate the second finite differences of that array (i.e.
# calculate the difference array and then calculate the difference
# array of the difference array)
d2 = np.diff(np.diff(probabilities))
# Get the absolute values of all those second finite differences
d2 = np.abs(d2)
# Normalize (all elements in the array are divided by the sum of the
# array's elements)
d2 = d2 / d2.sum(axis=-1, keepdims=True)
# Get the prefix-sum array
cumulative_d2 = np.cumsum(d2, axis=-1)
# We will remove the tokens with a cumulative normalized absolute
# second finite difference larger than the TFS value
sorted_indices_to_remove = cumulative_d2 > tfs
# Don't remove the token with the highest logit
sorted_indices_to_remove[0] = False
# Since the d2 array has two fewer elements than the logits array,
# we'll add two extra Trues to the end
sorted_indices_to_remove = np.pad(
sorted_indices_to_remove,
(0, 2),
constant_values=True,
)
# Unsort and remove
_, indices_to_remove = jax.lax.sort_key_val(
np.argsort(-logits),
sorted_indices_to_remove,
)
return np.where(indices_to_remove, -np.inf, logits)
# Typical sampling (https://arxiv.org/pdf/2202.00666.pdf)
def typical_filter(logits):
# Compute softmax probabilities and the natural logarithms of them
probs = jax.nn.softmax(logits)
with np.errstate(divide="ignore"):
log_probs = np.log(probs)
# Compute the negative of entropy, which is the sum of p*ln(p) for all p
# in the set of softmax probabilities of the logits
neg_entropy = np.nansum(probs * log_probs, axis=-1, keepdims=True)
# Determine absolute difference between the negative entropy and the
# log probabilities
entropy_deviation = np.abs(neg_entropy - log_probs)
# Keep certain tokens such that the sum of the entropy_deviation of the
# kept tokens is the smallest possible value such that the sum of the
# softmax probabilities of the kept tokens is at least the threshold
# value (by sorting the tokens in ascending order of entropy_deviation
# and then keeping the smallest possible number of tokens from the
# beginning such that sum of softmax probabilities is at or above the
# threshold)
_, sorted_logits = jax.lax.sort_key_val(entropy_deviation, probs)
sorted_indices_to_remove = np.cumsum(sorted_logits, axis=-1) >= typical
sorted_indices_to_remove = np.roll(sorted_indices_to_remove, 1, axis=-1)
sorted_indices_to_remove[0] = False
# Unsort and remove
_, indices_to_remove = jax.lax.sort_key_val(
jnp.argsort(entropy_deviation),
sorted_indices_to_remove,
)
return np.where(indices_to_remove, -jnp.inf, logits)
# Temperature (just divide the logits by the temperature)
def temp_filter(logits):
return logits / temp
for k in sampler_order:
if k == 0 and top_k > 0: logits = top_k_filter(logits)
if k == 1 and top_a > 0.0: logits = top_a_filter(logits)
if k == 2 and top_p < 1.0: logits = top_p_filter(logits)
if k == 3 and tfs < 1.0: logits = tail_free_filter(logits)
if k == 4 and typical < 1.0: logits = typical_filter(logits)
if k == 5 and temp != 1.0: logits = temp_filter(logits)
if k == 6 and rpargs[1] != 1.0: logits = apply_repetition_penalty_dynamic(logits, *rpargs)
# Finally, pick one token using the softmax thingy again (it gives
# an array whose elements sum to 1 so it can be used nicely as a
# probability distribution)
return jax.random.categorical(key, logits, -1).astype(np.uint32)
def apply_repetition_penalty_static(logits, tokens, repetition_penalty, generated_index, gen_length, rpslope, rprange):
'''
This gets called by generate_loop_fn to apply repetition penalty
to the 1D array logits using the provided 1D array of tokens to penalize
'''
rpslope = jnp.int32(rpslope)
rprange = jnp.int32(rprange)
clipped_rprange = jax.lax.cond(rprange > 0, lambda x: x, lambda x: tokens.shape[-1], rprange)
penalty_arange = jnp.roll(jnp.arange(tokens.shape[-1]) + (clipped_rprange - tokens.shape[-1]), generated_index, axis=-1)
# Make a new array with the same length as the tokens array but with
# each element replaced by the value at the corresponding index in the
# logits array; e.g.
# if logits is [77, 5, 3, 98] and tokens is [0, 1, 2, 3, 2, 3, 1],
# then penalty_logits will be [77, 5, 3, 98, 3, 98, 5]
penalty_logits = jnp.take(logits, tokens)
# Repetition penalty slope
def apply_slope(carry):
repetition_penalty, rprange = carry
_penalty = (penalty_arange/(rprange - 1)) * 2 - 1
_penalty = (rpslope * _penalty) / (1 + jnp.abs(_penalty) * (rpslope - 1))
_penalty = 1 + ((_penalty + 1) / 2) * (repetition_penalty - 1)
return _penalty
repetition_penalty = jax.lax.cond(
(rpslope != 0.0) & (rprange > 0), # Not a typo; do not use `and` here, it makes JAX crash
apply_slope,
lambda carry: jnp.full(tokens.shape, carry[0]),
(repetition_penalty, rprange),
)
# Divide positive values by repetition_penalty and multiply negative
# values by repetition_penalty (the academic publication that described
# this technique actually just only divided, but that would cause tokens
# with negative logits to become more likely, which is obviously wrong)
penalty_logits = jnp.where(
penalty_arange >= 0,
jnp.where(
penalty_logits > 0,
penalty_logits/repetition_penalty,
penalty_logits*repetition_penalty,
),
penalty_logits,
)
# Finally, put those penalized logit values back into their original
# positions in the logits array
return logits.at[tokens].set(penalty_logits)
def kobold_sample_static(key, logits, rpargs, sampler_order: Optional[np.ndarray] = None, top_p=0.9, temp=0.5, top_k=0, tfs=1.0, typical=1.0, top_a=0.0):
'''
This gets called by generate_loop_fn to apply a series of 6 filters
to the logits (top-k, then top-a, then top-p, then TFS, then typical, then temperature)
before picking one token using the modified logits
'''
# Top-k (keep only the k tokens with the highest logits and remove
# the rest, by setting their logits to negative infinity)
def top_k_filter(logits):
# After sorting the logits array in descending order,
# sorted_indices_to_remove is a 1D array that is True for tokens
# in the sorted logits array we want to remove and False for ones
# we want to keep, in this case the first top_k elements will be
# False and the rest will be True
sorted_indices_to_remove = jnp.arange(len(logits)) >= top_k
# Unsort the logits array back to its original configuration and
# remove tokens we need to remove
_, indices_to_remove = jax.lax.sort_key_val(
jnp.argsort(-logits),
sorted_indices_to_remove,
)
return jnp.where(indices_to_remove, -jnp.inf, logits)
# Top-a (remove all tokens that have softmax probability less than
# a*m^2 where m is the maximum softmax probability)
def top_a_filter(logits):
# Replace every element in the logits array
# with e (Euler's number) to the power of that element, and divide
# each element of the new array by the sum of the elements in the
# new array
probabilities = jax.nn.softmax(logits)
# Find the largest probability
probs_max = probabilities.max()
# Remove tokens
return jnp.where(probabilities < probs_max * probs_max * top_a, -jnp.inf, logits)
# Top-p (after sorting the remaining tokens again in descending order of
# logit, remove the ones that have cumulative softmax probability
# greater than p)
def top_p_filter(logits):
# Sort the logits array in descending order, replace every element
# with e (Euler's number) to the power of that element, and divide
# each element of the new array by the sum of the elements in the
# new array
sorted_logits = -jnp.sort(-logits)
probabilities = jax.nn.softmax(sorted_logits)
# Calculate cumulative_probabilities as the prefix-sum array of
# probabilities
cumulative_probabilities = jnp.cumsum(probabilities, axis=-1)
# We want to remove tokens with cumulative probability higher
# than top_p
sorted_indices_to_remove = cumulative_probabilities > top_p
# Don't ever remove the token with the highest logit, even if
# the probability is higher than top_p
sorted_indices_to_remove = sorted_indices_to_remove.at[0].set(False)
# Unsort and remove
_, indices_to_remove = jax.lax.sort_key_val(
jnp.argsort(-logits),
sorted_indices_to_remove,
)
return jnp.where(indices_to_remove, -jnp.inf, logits)
# Tail free sampling (basically top-p a second time on remaining tokens
# except it's the "cumulative normalized absolute second finite
# differences of the softmax probabilities" instead of just the
# cumulative softmax probabilities)
def tail_free_filter(logits):
# Sort in descending order
sorted_logits = -jnp.sort(-logits)
# Softmax again
probabilities = jax.nn.softmax(sorted_logits)
# Calculate the second finite differences of that array (i.e.
# calculate the difference array and then calculate the difference
# array of the difference array)
d2 = jnp.diff(jnp.diff(probabilities))
# Get the absolute values of all those second finite differences
d2 = jnp.abs(d2)
# Normalize (all elements in the array are divided by the sum of the
# array's elements)
d2 = d2 / d2.sum(axis=-1, keepdims=True)
# Get the prefix-sum array
cumulative_d2 = jnp.cumsum(d2, axis=-1)
# We will remove the tokens with a cumulative normalized absolute
# second finite difference larger than the TFS value
sorted_indices_to_remove = cumulative_d2 > tfs
# Don't remove the token with the highest logit
sorted_indices_to_remove = sorted_indices_to_remove.at[0].set(False)
# Since the d2 array has two fewer elements than the logits array,
# we'll add two extra Trues to the end
sorted_indices_to_remove = jnp.pad(
sorted_indices_to_remove,
(0, 2),
constant_values=True,
)
# Unsort and remove
_, indices_to_remove = jax.lax.sort_key_val(
jnp.argsort(-logits),
sorted_indices_to_remove,
)
return jnp.where(indices_to_remove, -jnp.inf, logits)
# Typical sampling (https://arxiv.org/pdf/2202.00666.pdf)
def typical_filter(logits):
# Compute softmax probabilities and the natural logarithms of them
probs = jax.nn.softmax(logits)
log_probs = jnp.log(probs)
# Compute the negative of entropy, which is the sum of p*ln(p) for all p
# in the set of softmax probabilities of the logits
neg_entropy = jnp.nansum(probs * log_probs, axis=-1, keepdims=True)
# Determine absolute difference between the negative entropy and the
# log probabilities
entropy_deviation = jnp.abs(neg_entropy - log_probs)
# Keep certain tokens such that the sum of the entropy_deviation of the
# kept tokens is the smallest possible value such that the sum of the
# softmax probabilities of the kept tokens is at least the threshold
# value (by sorting the tokens in ascending order of entropy_deviation
# and then keeping the smallest possible number of tokens from the
# beginning such that sum of softmax probabilities is at or above the
# threshold)
_, sorted_logits = jax.lax.sort_key_val(entropy_deviation, probs)
sorted_indices_to_remove = jnp.cumsum(sorted_logits, axis=-1) >= typical
sorted_indices_to_remove = jnp.roll(sorted_indices_to_remove, 1, axis=-1)
sorted_indices_to_remove = sorted_indices_to_remove.at[0].set(False)
# Unsort and remove
_, indices_to_remove = jax.lax.sort_key_val(
jnp.argsort(entropy_deviation),
sorted_indices_to_remove,
)
return jnp.where(indices_to_remove, -jnp.inf, logits)
# Temperature (just divide the logits by the temperature)
def temp_filter(logits):
return logits / temp
for k in sampler_order:
logits = jax.lax.cond(jnp.logical_and(k == 0, top_k > 0), top_k_filter, lambda x: x, logits)
logits = jax.lax.cond(jnp.logical_and(k == 1, top_a > 0.0), top_a_filter, lambda x: x, logits)
logits = jax.lax.cond(jnp.logical_and(k == 2, top_p < 1.0), top_p_filter, lambda x: x, logits)
logits = jax.lax.cond(jnp.logical_and(k == 3, tfs < 1.0), tail_free_filter, lambda x: x, logits)
logits = jax.lax.cond(jnp.logical_and(k == 4, typical < 1.0), typical_filter, lambda x: x, logits)
logits = jax.lax.cond(jnp.logical_and(k == 5, temp != 1.0), temp_filter, lambda x: x, logits)
logits = jax.lax.cond(jnp.logical_and(k == 6, rpargs[1] != 1.0), lambda x: apply_repetition_penalty_static(*x), lambda x: x[0], (logits, *rpargs))
# Finally, pick one token using the softmax thingy again (it gives
# an array whose elements sum to 1 so it can be used nicely as a
# probability distribution)
return jax.random.categorical(key, logits, -1).astype(jnp.uint32)
pad_token_id = 50256
def sample_func(data, key, numseqs_aux, badwords, repetition_penalty, generated_index, gen_length, rpslope, rprange, sampler_options):
numseqs = numseqs_aux.shape[0]
gi = data[0][1]
def sample_loop_fn(carry):
generated, generated_index, logits, _ = carry[0][0]
sample_key = carry[1]
# Get the pseudo-random number generator key that will
# be used by kobold_sample_dynamic to randomly pick a token
sample_key, new_key = jax.random.split(sample_key, num=2)
# Remove any tokens in the badwords list by setting
# their logits to negative infinity which effectively
# makes their probabilities of being chosen zero
logits[badwords] = -np.inf
# Use the sampler (kobold_sample_dynamic) to pick one token
# based on the logits array as a 0D uint32 array
# (higher logit means higher probability of being
# picked, non-linearly)
next_token = kobold_sample_dynamic(
sample_key,
logits,
(
generated,
repetition_penalty,
generated_index,
gen_length,
rpslope,
rprange,
),
**sampler_options,
)
# Remember what token was picked
generated[generated_index] = next_token
generated_index += 1
# Re-pack the current sample_loop_fn's state so we can
# get back the same variables the next time
carry[0][0] = [generated, generated_index, logits, next_token]
carry[0].append(carry[0].pop(0))
return carry[0], new_key
# return jax.lax.while_loop(
# lambda carry: carry[0][0][1] == gi,
# sample_loop_fn,
# (data, key),
# )
carry = (data, key)
while carry[0][0][1] == gi:
carry = sample_loop_fn(carry)
return carry
class PenalizingCausalTransformer(CausalTransformer):
def __init__(self, config, **kwargs):
# Initialize
super().__init__(config, **kwargs)
def generate_static(state, key, ctx, ctx_length, gen_length, numseqs_aux, sampler_options, soft_embeddings=None):
compiling_callback()
numseqs = numseqs_aux.shape[0]
# These are the tokens that we don't want the AI to ever write
badwords = jnp.array(vars.badwordsids).squeeze()
@hk.transform
def generate_sample(context, ctx_length):
# Give the initial context to the transformer
transformer = CausalTransformerShard(config)
def generate_initial_scan_fn(sequence_index, _):
_, initial_state = transformer.generate_initial(context, ctx_length, soft_embeddings=soft_embeddings)
# The "generated" array will contain the tokens from the
# context as well as the tokens picked by the sampler at
# each stage, padded with a bunch of 50256s, so we know
# which tokens have to be repetition penalized
generated = jnp.pad(context, (0, config["seq"]), constant_values=pad_token_id) # Let it start off with just the 2048 context tokens, plus some 50256s which will be eventually filled with sampler-chosen tokens
generated_index = config["seq"]
# Add that information to generate_loop_fn's starting state
initial_state = (generated, generated_index, sequence_index) + initial_state
return sequence_index+1, initial_state
_, initial_states = jax.lax.scan(generate_initial_scan_fn, 0, None, numseqs)
sample_key = initial_states[-1][0]
initial_states = list(jax.tree_map(lambda x: x[i], initial_states[:-1]) for i in range(numseqs))
# Get repetition penalty from the arguments
repetition_penalty = sampler_options.pop('repetition_penalty', None)
rpslope = sampler_options.pop('rpslope', None)
rprange = sampler_options.pop('rprange', None)
# This is the main generation loop
def generate_loop_fn(carry):
# Unpack current generate_loop_fn state
generated, generated_index, sequence_index, next_token, decode_state = carry[0][0]
sample_key = carry[1]
# Get the pseudo-random number generator key that will
# be used by kobold_sample_static to randomly pick a token
sample_key, new_key = jax.random.split(sample_key)
# Give the context to the model and get the logits it
# spits out
# (a 2D array with 1 row and 50400 columns representing
# how strongly it thinks each of the 50257 tokens in its
# vocabulary should be appended to the context, followed
# by 143 apparently useless columns ???)
logits, new_state = transformer.generate_once(next_token, decode_state, soft_embeddings=soft_embeddings)
# Verify that logits does indeed have that many rows and
# columns (if you get an error here, pray for mercy)
assert logits.shape == (1, config["n_vocab"])
# Flatten it into a 1D array to make it easier to use
logits = logits[0]
# Remove any tokens in the badwords list by setting
# their logits to negative infinity which effectively
# makes their probabilities of being chosen zero
logits = logits.at[badwords].set(-jnp.inf)
# Use the sampler (kobold_sample_static) to pick one token
# based on the logits array as a 0D uint32 array
# (higher logit means higher probability of being
# picked, non-linearly)
next_token = kobold_sample_static(
sample_key,
logits,
(
generated,
repetition_penalty,
generated_index,
gen_length,
rpslope,
rprange,
),
**sampler_options,
)
# Remember what token was picked
generated = generated.at[generated_index].set(next_token)
generated_index += 1
# Re-pack the current generate_loop_fn's state so we can
# get back the same variables the next time
carry[0][0] = (generated, generated_index, sequence_index, next_token[jnp.newaxis], new_state)
carry[0].append(carry[0].pop(0))
return carry[0], new_key
return jax.lax.while_loop(
lambda carry: carry[0][0][1] - config["seq"] < gen_length,
generate_loop_fn,
(initial_states, sample_key),
)
return generate_sample.apply(state["params"], key, ctx, ctx_length)
self.generate_static_xmap = jax.experimental.maps.xmap(
fun=generate_static,
in_axes=(
["shard", ...],
["batch", ...],
["batch", ...],
["batch", ...],
["batch", ...],
["batch", ...],
["batch", ...],
["shard", ...],
),
out_axes=["shard", "batch", ...],
axis_resources={'shard': 'mp', 'batch': 'dp'},
)
def generate_initial(state, key, ctx, ctx_length, numseqs_aux, soft_embeddings=None):
compiling_callback()
numseqs = numseqs_aux.shape[0]
@hk.transform
def generate_initial_inner(context, ctx_length):
# Give the initial context to the transformer
transformer = CausalTransformerShard(config)
def generate_initial_scan_fn(sequence_index, c):
_, initial_state = transformer.generate_initial(c, ctx_length, soft_embeddings=soft_embeddings)
generated_index = config["seq"]
# Add that information to generate_loop_fn's starting state
initial_state = (jnp.empty(config["n_vocab"], dtype=jnp.float32), generated_index, sequence_index) + initial_state
return sequence_index+1, initial_state
_, initial_states = jax.lax.scan(generate_initial_scan_fn, 0, context, numseqs)
sample_key = initial_states[-1][0]
initial_states = list(list(jax.tree_map(lambda x: x[i], initial_states[:-1])) for i in range(numseqs))
return initial_states, sample_key
return generate_initial_inner.apply(state["params"], key, ctx, ctx_length)
self.generate_initial_xmap = jax.experimental.maps.xmap(
fun=generate_initial,
in_axes=(
["shard", ...],
["batch", ...],
["batch", ...],
["batch", ...],
["batch", ...],
["shard", ...],
),
out_axes=["shard", "batch", ...],
axis_resources={'shard': 'mp', 'batch': 'dp'},
)
def generate_once(data, state, numseqs_aux, soft_embeddings=None):
numseqs = numseqs_aux.shape[0]
@hk.without_apply_rng
@hk.transform
def generate_once_inner():
gi = data[0][1]
# Give the initial context to the transformer
transformer = CausalTransformerShard(config)
# This is the main generation loop
def generate_loop_fn(carry):
# Unpack current generate_loop_fn state
_, generated_index, sequence_index, next_token, decode_state = carry[0][0]
# Give the context to the model and get the logits it
# spits out
# (a 2D array with 1 row and 50400 columns representing
# how strongly it thinks each of the 50257 tokens in its
# vocabulary should be appended to the context, followed
# by 143 apparently useless columns ???)
logits, new_state = transformer.generate_once(next_token, decode_state, soft_embeddings=soft_embeddings)
# Verify that logits does indeed have that many rows and
# columns (if you get an error here, pray for mercy)
assert logits.shape == (1, config["n_vocab"])
assert logits.dtype == jnp.float32
# Flatten it into a 1D array to make it easier to use
logits = logits[0]
# Re-pack the current generate_loop_fn's state so we can
# get back the same variables the next time
generated_index += 1
carry[0][0] = [logits, generated_index, sequence_index, next_token, new_state]
carry[0].append(carry[0].pop(0))
return carry[0],
return jax.lax.while_loop(
lambda carry: carry[0][0][1] == gi,
generate_loop_fn,
(data,),
)
return generate_once_inner.apply(state["params"])
self.generate_once_xmap = jax.experimental.maps.xmap(
fun=generate_once,
in_axes=(
["shard", "batch", ...],
["shard", ...],
["batch", ...],
["shard", ...],
),
out_axes=["shard", "batch", ...],
axis_resources={'shard': 'mp', 'batch': 'dp'},
)
def generate_dynamic(self, ctx, ctx_length, gen_length, numseqs, return_logits=False, soft_embeddings=None, excluded_world_info=None, use_callback=True):
assert excluded_world_info is not None
assert not return_logits
assert gen_length.ndim == 1
assert soft_embeddings is not None
key = hk.PRNGSequence(rng.randint(0, 2 ** 60))
batch_size = ctx.shape[0]
self.batch_size = batch_size
_numseqs_aux = jnp.empty((batch_size, numseqs), dtype=np.uint32)
numseqs_aux = batch_xmap(_numseqs_aux)
sample_data = [
[
np.pad(ctx[0][i], (0, params["seq"]), constant_values=pad_token_id),
params["seq"],
None,
np.empty((), dtype=np.uint32),
]
for i in range(numseqs)
]
n_generated = 0
regeneration_required = False
halt = False
started_compiling_callback()
generate_data, sample_key = self.generate_initial_xmap(self.state, jnp.array(key.take(batch_size)), ctx, ctx_length, numseqs_aux, soft_embeddings)
sample_key = np.asarray(sample_key[0, 0])
while True:
generate_data, = self.generate_once_xmap(generate_data, self.state, numseqs_aux, soft_embeddings)
for i in range(numseqs):
sample_data[i][2] = np.array(generate_data[i][0][0, 0], copy=True)
if use_callback:
logits = np.float32(tuple(d[2] for d in sample_data))
logits = warper_callback(logits)
for i in range(numseqs):
sample_data[i][2] = logits[i]
sampler_options = settings_callback()
repetition_penalty = sampler_options.pop("repetition_penalty", 1.0)
rpslope = sampler_options.pop("rpslope", 0.0)
rprange = sampler_options.pop("rprange", 0)
sample_data, sample_key = sample_func(sample_data, sample_key, _numseqs_aux, badwords, repetition_penalty, params["seq"] + n_generated, gen_length, rpslope, rprange, sampler_options)
n_generated += 1
for i in range(numseqs):
generate_data[i][3] = np.tile(sample_data[i][0][sample_data[i][1]-1][np.newaxis, np.newaxis], (params["cores_per_replica"], 1, 1))
if use_callback:
generated = np.uint32(tuple(d[0] for d in sample_data))
excluded_world_info, regeneration_required, halt = stopping_callback(generated, n_generated, excluded_world_info)
if regeneration_required or halt:
break
else:
break
stopped_compiling_callback()
return sample_data, n_generated, regeneration_required, halt
def generate_static(self, ctx, ctx_length, gen_length, numseqs, sampler_options, return_logits=False, soft_embeddings=None):
assert not return_logits
key = hk.PRNGSequence(rng.randint(0, 2 ** 60))
batch_size = ctx.shape[0]
self.batch_size = batch_size
started_compiling_callback()
result = self.generate_static_xmap(
self.state,
jnp.array(key.take(batch_size)),
ctx,
np.array(ctx_length, dtype=np.uint32),
np.array(gen_length, dtype=np.uint32),
np.empty((batch_size, numseqs), dtype=np.uint8),
sampler_options,
soft_embeddings,
)
stopped_compiling_callback()
return result
def infer_dynamic(
context: np.array,
numseqs=1,
gen_len=80,
soft_embeddings: Optional[np.array] = None,
soft_tokens: Optional[np.array] = None,
excluded_world_info = None,
use_callback=True,
) -> Tuple[List[np.array], int, bool, bool]:
assert excluded_world_info is not None
maps.thread_resources.env = thread_resources_env
total_batch = 1
tokens = context
if(soft_tokens is not None):
tokens = np.uint32(np.concatenate((np.tile(soft_tokens, (tokens.shape[0], 1)), tokens), axis=-1))
provided_ctx = tokens.shape[-1]
pad_amount = seq - provided_ctx
padded_tokens = np.pad(tokens, ((0, 0), (pad_amount, 0)), constant_values=pad_token_id)
batched_tokens = np.array([padded_tokens] * total_batch)
samples = []
output = network.generate_dynamic(
batched_tokens,
np.ones(total_batch, dtype=np.uint32) * provided_ctx,
np.ones(total_batch, dtype=np.uint32) * gen_len,
numseqs,
soft_embeddings=soft_embeddings,
excluded_world_info=excluded_world_info,
use_callback=use_callback,
)
for out in output[0]:
samples.append(out[0][params["seq"] : params["seq"] + gen_len])
return (samples,) + output[1:]
def infer_static(
context: np.array,
top_p=0.9,
temp=0.5,
top_k=0,
tfs=1.0,
typical=1.0,
top_a=0.0,
repetition_penalty=1.0,
rpslope=0.0,
rprange=0,
numseqs=1,
gen_len=80,
soft_embeddings: Optional[np.array] = None,
soft_tokens: Optional[np.array] = None,
sampler_order: Optional[List[int]] = None,
) -> List[np.array]:
maps.thread_resources.env = thread_resources_env
if sampler_order is None:
sampler_order = utils.default_sampler_order.copy()
sampler_order = sampler_order[:]
if len(sampler_order) < 7: # Add repetition penalty at beginning if it's not present
sampler_order = [6] + sampler_order
sampler_order = np.uint32(sampler_order)
total_batch = 1
tokens = context
if(soft_tokens is not None):
tokens = np.uint32(np.concatenate((soft_tokens, tokens)))
provided_ctx = tokens.shape[0]
pad_amount = seq - provided_ctx
padded_tokens = np.pad(tokens, ((pad_amount, 0),), constant_values=pad_token_id)
batched_tokens = np.array([padded_tokens] * total_batch)
samples = []
batched_generator_params = {
"sampler_order": np.repeat(sampler_order[np.newaxis], total_batch, axis=0),
"temp": temp * np.ones(total_batch),
"top_p": top_p * np.ones(total_batch),
"tfs": tfs * np.ones(total_batch),
"typical": typical * np.ones(total_batch),
"top_a": top_a * np.ones(total_batch),
"repetition_penalty": repetition_penalty * np.ones(total_batch),
"rpslope": rpslope * np.ones(total_batch),
"rprange": np.full(total_batch, rprange, dtype=np.uint32),
"top_k": np.full(total_batch, top_k, dtype=np.uint32)
}
output = network.generate_static(
batched_tokens,
np.ones(total_batch, dtype=np.uint32) * provided_ctx,
np.ones(total_batch, dtype=np.uint32) * gen_len,
numseqs,
batched_generator_params,
soft_embeddings=soft_embeddings,
)[0]
for o in output:
samples.append(o[0][0, 0, params["seq"] : params["seq"] + gen_len])
return samples
def reshard_reverse(x, total_shards, old_shape):
assert len(x.shape) != 1
if len(x.shape) == 2:
if old_shape[1] == x.shape[1]:
out = x[0:1].tile((total_shards, 1))
else:
out = x.reshape(old_shape)
elif len(x.shape) == 3:
if x.shape[0] * x.shape[2] == old_shape[2]:
out = x.reshape(old_shape)
elif x.shape[0] * x.shape[1] == old_shape[1]:
out = x.reshape((old_shape[1], old_shape[0], old_shape[2])).permute((1, 0, 2))
else:
assert False
else:
assert False
return out
def get_old_shape(t, total_shards, dim=2):
if len(t.shape) == 2:
shard_shape = t.shape
if dim == 1:
assert shard_shape[0] % total_shards == 0
return (shard_shape[0] // total_shards, shard_shape[1])
elif dim == 2:
assert shard_shape[1] % total_shards == 0
return (shard_shape[0], shard_shape[1] // total_shards)
else:
raise ValueError(f"Unsupported dim {dim}")
if len(t.shape) == 1:
assert t.shape[0] % total_shards == 0
return (t.shape[0] // total_shards,)
else:
raise ValueError(f"Unsupported shape {t.shape}")
def read_neox_checkpoint(state, path, config, checkpoint_shards=2):
assert config["cores_per_replica"] % checkpoint_shards == 0
output_shards = config["cores_per_replica"] // checkpoint_shards
import torch
import torch.utils.dlpack
import torch_lazy_loader
from tqdm.auto import tqdm
move_xmap = jax.experimental.maps.xmap(
fun=lambda x, _: to_bf16(x),
in_axes=(["shard", ...], ["batch", ...]),
out_axes=["shard", ...],
axis_resources={'shard': 'mp', 'batch': 'dp'}
)
path_template = os.path.join(path, "layer_{layer:02d}-model_{shard:02d}-model_states.pt")
static_mapping = {
"word_embeddings.weight": {"module": "embedding_shard/~/linear", "param": "w", "axis": 1},
"final_linear.weight": {"module": "projection_shard/~/linear", "param": "w", "axis": 2},
"norm.weight": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale", "axis": None},
"norm.bias": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset", "axis": None},
}
layer_mapping = {
"attention.query_key_value.weight": {"module": "combined_qkv", "param": "w", "axis": 2},
"attention.query_key_value.bias": {"module": "combined_qkv", "param": "b", "axis": 1},
"attention.dense.weight": {"module": "linear_3", "param": "w", "axis": 1},
"attention.dense.bias": {"module": "linear_3", "param": "b", "axis": None},
"mlp.dense_h_to_4h.weight": {"module": "linear_4", "param": "w", "axis": 2},
"mlp.dense_h_to_4h.bias": {"module": "linear_4", "param": "b", "axis": 1},
"mlp.dense_4h_to_h.weight": {"module": "linear_5", "param": "w", "axis": 1},
"mlp.dense_4h_to_h.bias": {"module": "linear_5", "param": "b", "axis": None},
"input_layernorm.weight": {"module": "replicated_layer_norm", "param": "scale", "axis": None},
"input_layernorm.bias": {"module": "replicated_layer_norm", "param": "offset", "axis": None},
"post_attention_layernorm.weight": {"module": "replicated_layer_norm_1", "param": "scale", "axis": None},
"post_attention_layernorm.bias": {"module": "replicated_layer_norm_1", "param": "offset", "axis": None},
}
tqdm_length = len(static_mapping) + config["layers"]*len(layer_mapping)
bar = tqdm(total=tqdm_length, desc="Loading from NeoX checkpoint")
for checkpoint_layer in range(config["layers"] + 5):
if checkpoint_layer in (1, config["layers"] + 2):
continue
layer = checkpoint_layer - 2
shards = []
with torch_lazy_loader.use_custom_unpickler(torch_lazy_loader.RestrictedUnpickler):
for checkpoint_shard in range(checkpoint_shards):
shards.append(torch.load(path_template.format(layer=checkpoint_layer, shard=checkpoint_shard), map_location="cpu"))
for key in shards[0]:
if key == "attention.rotary_emb.inv_freq":
continue
elif key in static_mapping:
target_module = "causal_transformer_shard/~/" + static_mapping[key]["module"]
target_param = static_mapping[key]["param"]
target_axis = static_mapping[key]["axis"]
elif key in layer_mapping:
target_module = f"causal_transformer_shard/~/layer_{layer}/~/" + layer_mapping[key]["module"]
target_param = layer_mapping[key]["param"]
target_axis = layer_mapping[key]["axis"]
else:
error = f"{repr(key)} not found in mapping"
print("\n\nERROR: ", error, file=sys.stderr)
raise RuntimeError(error)
original_shape = shards[0][key].shape
for checkpoint_shard in range(checkpoint_shards):
if key in ("attention.dense.bias", "mlp.dense_4h_to_h.bias"):
shards[checkpoint_shard][key] /= output_shards
if key != "word_embeddings.weight" and shards[checkpoint_shard][key].ndim == 2:
shards[checkpoint_shard][key] = shards[checkpoint_shard][key].T
tensor = shards[checkpoint_shard][key]
if target_axis is not None:
target_shape = (output_shards,) + get_old_shape(tensor, total_shards=output_shards, dim=target_axis)
else:
target_shape = (output_shards, tensor.shape[0])
shards[checkpoint_shard][key] = reshard_reverse(tensor.unsqueeze_(0), output_shards, target_shape)
#print(key, ":", original_shape, "->", shards[0][key].shape)
tensor = torch.cat([shards[s][key] for s in range(checkpoint_shards)], dim=0)
target_shape = state["params"][target_module][target_param].shape
if tensor.shape != target_shape:
error = f"Weight {repr(key)} has shape {tensor.shape} in checkpoint but shape {target_shape} was requested by MTJ for {target_module} {target_param}"
print("\n\nERROR: ", error, file=sys.stderr)
raise RuntimeError(error)
if tensor.dtype is torch.float16 or tensor.dtype is torch.float32:
tensor = tensor.bfloat16()
state["params"][target_module][target_param] = move_xmap(
jax.dlpack.from_dlpack(torch.utils.dlpack.to_dlpack(tensor)).copy(),
np.zeros(config["cores_per_replica"]),
)
bar.update(1)
for mk, mv in state["params"].items():
for pk, pv in mv.items():
if isinstance(pv, PlaceholderTensor):
error = f"{mk} {pk} could not be found in the model checkpoint"
print("\n\nERROR: " + error, file=sys.stderr)
raise RuntimeError(error)
def load_model(path: str, driver_version="tpu_driver_20221109", hf_checkpoint=False, socketio_queue=None, initial_load=False, logger=None, **kwargs) -> None:
global thread_resources_env, seq, tokenizer, network, params, pad_token_id
if "pad_token_id" in kwargs:
pad_token_id = kwargs["pad_token_id"]
elif "eos_token_id" in kwargs:
pad_token_id = kwargs["eos_token_id"]
if not hasattr(vars, "sampler_order") or not vars.sampler_order:
vars.sampler_order = utils.default_sampler_order.copy()
default_params = {
"compat": "j",
"layers": 28,
"d_model": 4096,
"n_heads": 16,
"n_vocab": 50400,
"n_vocab_padding": 0,
"norm": "layernorm",
"pe": "rotary",
"pe_rotary_dims": 64,
"seq": 2048,
"cores_per_replica": 8,
"tokenizer_class": "GPT2Tokenizer",
"tokenizer": "gpt2",
}
params = kwargs
if vars.model == "TPUMeshTransformerGPTNeoX":
default_params = {
"compat": "neox",
"layers": 44,
"d_model": 6144,
"n_heads": 64,
"n_vocab": 50432,
"n_vocab_padding": 0,
"norm": "doublelayernorm",
"pe": "neox_rotary",
"pe_rotary_dims": 24,
"seq": 2048,
"cores_per_replica": 8,
"tokenizer_class": "GPT2Tokenizer",
"tokenizer": "gpt2",
}
# Try to convert HF config.json to MTJ config
if hf_checkpoint:
spec_path = os.path.join("maps", vars.model_type + ".json")
if not os.path.isfile(spec_path):
raise NotImplementedError(f"Unsupported model type {repr(vars.model_type)}")
with open(spec_path) as f:
lazy_load_spec = json.load(f)
if "mtj_compat" in lazy_load_spec:
params["compat"] = lazy_load_spec["mtj_compat"]
if "mtj_pe" in lazy_load_spec:
params["pe"] = lazy_load_spec["mtj_pe"]
for k, v in lazy_load_spec.get("mtj_config_map", {}).items():
if type(v) is not list:
params[k] = params[v]
continue
for i in range(len(v)):
if i == len(v) - 1:
params[k] = v[i]
elif v[i] in params:
params[k] = params[v[i]]
break
params["n_vocab"] = params["vocab_size"]
if "activation_function" in params:
params["activation"] = params["activation_function"]
# Both the number of attention heads in the model and the embedding
# dimension of the model need to be divisible by the number of TPU cores
# that we use, and JAX also requires the number of TPU cores used to be
# an even number if we're using more than one core, so logically we try
# to pick the largest possible even number of TPU cores such that the
# number of attention heads and embedding dimension are both divisible
# by the number of TPU cores, and fall back to one core if an even
# number of TPU cores is not possible.
for c in (8, 6, 4, 2, 1):
if 0 == params["n_heads"] % c == params.get("d_embed", params["d_model"]) % c:
params["cores_per_replica"] = c
break
# The vocabulary size of the model also has to be divisible by the
# number of TPU cores, so we pad the vocabulary with the minimum
# possible number of dummy tokens such that it's divisible.
params["n_vocab_padding"] = -(params["n_vocab"] % -params["cores_per_replica"])
if "compat" in params:
default_params["compat"] = params["compat"]
if default_params["compat"] == "fairseq_lm":
default_params["tokenizer"] = "KoboldAI/fairseq-dense-125M"
for param in default_params:
if param not in params:
params[param] = default_params[param]
# Use an optimization that will allow us to avoid one extra transpose operation
if hf_checkpoint:
params["transposed_linear"] = True
# Load tokenizer
if vars.model == "TPUMeshTransformerGPTNeoX":
tokenizer = Tokenizer.from_file(os.path.join(path, "20B_tokenizer.json"))
def new_encode(old_encode):
def encode(s, *args, **kwargs):
return old_encode(s).ids
return encode
tokenizer.encode = new_encode(tokenizer.encode)
tokenizer._koboldai_header = []
elif not hf_checkpoint:
if not isinstance(params["tokenizer_class"], str) or not any(params["tokenizer_class"].endswith(s) for s in ("Tokenizer", "TokenizerFast")):
raise ValueError("`tokenizer_class` must be a string ending in 'Tokenizer' or 'TokenizerFast'")
tokenizer_class = getattr(__import__("transformers"), params["tokenizer_class"])
tokenizer = tokenizer_class.from_pretrained(params["tokenizer"])
# Disable JAX warnings about these two functions having been renamed
jax.host_count = jax.process_count
jax.host_id = jax.process_index
print("Connecting to your Colab instance's TPU", flush=True)
spinner = multiprocessing.Process(target=show_spinner, args=())
spinner.start()
if os.environ.get('COLAB_TPU_ADDR', '') != '':
tpu_address = os.environ['COLAB_TPU_ADDR'] # Colab
else:
tpu_address = os.environ['TPU_NAME'] # Kaggle
tpu_address = tpu_address.replace("grpc://", "")
tpu_address_without_port = tpu_address.split(':', 1)[0]
url = f'http://{tpu_address_without_port}:8475/requestversion/{driver_version}'
requests.post(url)
config.FLAGS.jax_xla_backend = "tpu_driver"
config.FLAGS.jax_backend_target = "grpc://" + tpu_address
spinner.terminate()
print()
cores_per_replica = params["cores_per_replica"]
seq = params["seq"]
params["optimizer"] = _DummyOptimizer()
mesh_shape = (1, cores_per_replica)
devices = np.array(jax.devices()[:cores_per_replica]).reshape(mesh_shape)
thread_resources_env = maps.ResourceEnv(maps.Mesh(devices, ('dp', 'mp')), ())
maps.thread_resources.env = thread_resources_env
global badwords
# These are the tokens that we don't want the AI to ever write
badwords = jnp.array(vars.badwordsids).squeeze()
if not path.endswith("/"):
path += "/"
network = PenalizingCausalTransformer(params, dematerialized=True)
if not hf_checkpoint and vars.model != "TPUMeshTransformerGPTNeoX":
network.state = read_ckpt_lowmem(network.state, path, devices.shape[1])
#network.state = network.move_xmap(network.state, np.zeros(cores_per_replica))
return
if vars.model == "TPUMeshTransformerGPTNeoX":
print("\n\n\nThis model has ", f"{hk.data_structures.tree_size(network.state['params']):,d}".replace(",", " "), " parameters.\n")
read_neox_checkpoint(network.state, path, params)
return
# Convert from HF checkpoint
move_xmap = jax.experimental.maps.xmap(
fun=lambda x, _: to_bf16(x),
in_axes=(["shard", ...], ["batch", ...]),
out_axes=["shard", ...],
axis_resources={'shard': 'mp', 'batch': 'dp'}
)
model_spec = {}
for key, spec in lazy_load_spec.get("static_weights", {}).items():
if spec.get("mtj") is not None:
model_spec[key] = spec["mtj"].copy()
model_spec[key]["module"] = "causal_transformer_shard/~/" + model_spec[key]["module"]
for _key, spec in lazy_load_spec.get("layer_weights", {}).items():
for layer in range(params["layers"]):
if spec.get("mtj") is not None:
key = _key.format(layer=layer)
model_spec[key] = spec["mtj"].copy()
model_spec[key]["module"] = "causal_transformer_shard/~/" + model_spec[key]["module"].format(layer=layer)
import torch_lazy_loader
import torch
from tqdm.auto import tqdm
import functools
def callback(model_dict, f, **_):
if callback.nested:
return
callback.nested = True
with zipfile.ZipFile(f, "r") as z:
try:
last_storage_key = None
zipfolder = os.path.basename(os.path.normpath(f)).split('.')[0]
f = None
current_offset = 0
if utils.current_shard == 0:
print("\n\n\nThis model has ", f"{hk.data_structures.tree_size(network.state['params']):,d}".replace(",", " "), " parameters.\n")
if utils.num_shards is None or utils.current_shard == 0:
if utils.num_shards is not None:
num_tensors = len(utils.get_sharded_checkpoint_num_tensors(utils.from_pretrained_model_name, utils.from_pretrained_index_filename, **utils.from_pretrained_kwargs))
else:
num_tensors = len(model_dict)
utils.bar = tqdm(total=num_tensors, desc="Loading model tensors")
if utils.num_shards is not None:
utils.current_shard += 1
for key in sorted(model_dict.keys(), key=lambda k: (model_dict[k].key, model_dict[k].seek_offset)):
model_spec_key = max((k for k in model_spec.keys() if key.endswith(k)), key=len, default=None)
# Some model weights are used by transformers but not by MTJ.
# We have to materialize these weights anyways because
# transformers will throw a tantrum otherwise. To attain
# the least possible memory usage, we create them as meta
# tensors, which don't take up any actual CPU or TPU memory.
if model_spec_key is None:
model_dict[key] = torch.empty(model_dict[key].shape, dtype=model_dict[key].dtype, device="meta")
utils.bar.update(1)
continue
storage_key = model_dict[key].key
if storage_key != last_storage_key or model_dict[key].seek_offset < current_offset:
last_storage_key = storage_key
if isinstance(f, zipfile.ZipExtFile):
f.close()
try:
f = z.open(f"archive/data/{storage_key}")
except:
f = z.open(f"{zipfolder}/data/{storage_key}")
current_offset = 0
if current_offset != model_dict[key].seek_offset:
f.read(model_dict[key].seek_offset - current_offset)
current_offset = model_dict[key].seek_offset
spec = model_spec[model_spec_key]
transforms = set(spec.get("transforms", ()))
if not isinstance(model_dict[key], torch_lazy_loader.LazyTensor):
error = f"Duplicate key {repr(key)}"
print("\n\nERROR: " + error, file=sys.stderr)
raise RuntimeError(error)
size = functools.reduce(lambda x, y: x * y, model_dict[key].shape, 1)
dtype = model_dict[key].dtype
nbytes = size if dtype is torch.bool else size * ((torch.finfo if dtype.is_floating_point else torch.iinfo)(dtype).bits >> 3)
tensor = model_dict[key].materialize(f, map_location="cpu")
model_dict[key] = tensor.to("meta")
current_offset += nbytes
# MTJ requires certain mathematical operations to be performed
# on tensors in order for them to be in the correct format
if "remove_first_two_rows" in transforms:
tensor = tensor[2:]
if "divide_by_shards" in transforms:
tensor /= params["cores_per_replica"]
if "vocab_pad" in transforms:
tensor = torch.nn.functional.pad(tensor, (0,) * (tensor.ndim * 2 - 1) + (params["n_vocab_padding"],))
# We don't need to transpose linear module weights anymore because MTJ will do it for us if `transposed_linear` is set to True in the config
#if "no_transpose" not in transforms and tensor.ndim == 2:
# tensor = tensor.T
tensor.unsqueeze_(0)
# Shard the tensor so that parts of the tensor can be used
# on different TPU cores
tensor = reshard_reverse(
tensor,
params["cores_per_replica"],
network.state["params"][spec["module"]][spec["param"]].shape,
)
tensor = jnp.array(tensor.detach())
if tensor.dtype is torch.float16 or tensor.dtype is torch.float32:
tensor = tensor.bfloat16()
network.state["params"][spec["module"]][spec["param"]] = move_xmap(
tensor,
np.empty(params["cores_per_replica"]),
)
utils.bar.update(1)
if utils.num_shards is not None and utils.current_shard < utils.num_shards:
return
# Check for tensors that MTJ needs that were not provided in the
# HF model
for mk, mv in network.state["params"].items():
for pk, pv in mv.items():
if isinstance(pv, PlaceholderTensor):
# The transformers GPT-J models apparently do not
# have embedding bias, whereas MTJ GPT-J models do,
# so we have to supplement an embedding bias tensor
# by creating a tensor with the necessary shape, filled
# with zeros.
if mk == "causal_transformer_shard/~/embedding_shard/~/linear" and pk == "b":
mv[pk] = move_xmap(jnp.zeros(mv[pk].shape, dtype=jnp.bfloat16), np.empty(params["cores_per_replica"]))
else:
error = f"{mk} {pk} could not be found in the model checkpoint"
print("\n\nERROR: " + error, file=sys.stderr)
raise RuntimeError(error)
finally:
if utils.num_shards is None or utils.current_shard >= utils.num_shards:
utils.bar.close()
utils.bar = None
callback.nested = False
if isinstance(f, zipfile.ZipExtFile):
f.close()
callback.nested = False
if os.path.isdir(vars.model.replace('/', '_')):
import shutil
shutil.move(vars.model.replace('/', '_'), "models/{}".format(vars.model.replace('/', '_')))
print("\n", flush=True)
with torch_lazy_loader.use_lazy_torch_load(callback=callback, dematerialized_modules=True):
if(os.path.isdir(vars.custmodpth)):
try:
tokenizer = AutoTokenizer.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache", use_fast=False)
except Exception as e:
try:
tokenizer = AutoTokenizer.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
except Exception as e:
try:
tokenizer = GPT2Tokenizer.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
except Exception as e:
tokenizer = GPT2Tokenizer.from_pretrained("gpt2", revision=vars.revision, cache_dir="cache")
try:
model = AutoModelForCausalLM.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
except Exception as e:
model = GPTNeoForCausalLM.from_pretrained(vars.custmodpth, revision=vars.revision, cache_dir="cache")
elif(os.path.isdir("models/{}".format(vars.model.replace('/', '_')))):
try:
tokenizer = AutoTokenizer.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache", use_fast=False)
except Exception as e:
try:
tokenizer = AutoTokenizer.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
except Exception as e:
try:
tokenizer = GPT2Tokenizer.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
except Exception as e:
tokenizer = GPT2Tokenizer.from_pretrained("gpt2", revision=vars.revision, cache_dir="cache")
try:
model = AutoModelForCausalLM.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
except Exception as e:
model = GPTNeoForCausalLM.from_pretrained("models/{}".format(vars.model.replace('/', '_')), revision=vars.revision, cache_dir="cache")
else:
try:
tokenizer = AutoTokenizer.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache", use_fast=False)
except Exception as e:
try:
tokenizer = AutoTokenizer.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
except Exception as e:
try:
tokenizer = GPT2Tokenizer.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
except Exception as e:
tokenizer = GPT2Tokenizer.from_pretrained("gpt2", revision=vars.revision, cache_dir="cache")
try:
model = AutoModelForCausalLM.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
except Exception as e:
model = GPTNeoForCausalLM.from_pretrained(vars.model, revision=vars.revision, cache_dir="cache")
#network.state = network.move_xmap(network.state, np.zeros(cores_per_replica))
global shard_xmap, batch_xmap
shard_xmap = __shard_xmap()
batch_xmap = __batch_xmap(shard_dim=cores_per_replica)