|
import os |
|
import pickle |
|
import torch |
|
from mamba_lm import MambaLMConfig, from_pretrained |
|
from mamba_ssm import MambaLMHeadModel |
|
from contextlib import nullcontext |
|
import numpy as np |
|
from functools import partial |
|
import chess |
|
from sklearn.linear_model import LinearRegression |
|
import torch.nn as nn |
|
import torch.optim as optim |
|
import wandb |
|
import math |
|
import json |
|
|
|
BASE_DIR = "mamba/" |
|
|
|
class MambaPlayer: |
|
def __init__(self, model_name: str, move_num_in_gamestate: bool=False, update_contrastive: bool=False, update_linear: bool=False, linear_probe_path: str=None): |
|
self.model_name = model_name |
|
self.move_num_in_gamestate = move_num_in_gamestate |
|
|
|
|
|
init_from = "resume" |
|
out_dir = "out" |
|
device = "cuda" if torch.cuda.is_available() else "cpu" |
|
|
|
dtype = 'bfloat16' if torch.cuda.is_bf16_supported() else 'float32' |
|
seed = 1337 |
|
compile = False |
|
|
|
|
|
torch.manual_seed(seed) |
|
torch.cuda.manual_seed(seed) |
|
|
|
device_type = ( |
|
"cuda" if "cuda" in device else "cpu" |
|
) |
|
ptdtype = { |
|
"float32": torch.float32, |
|
"bfloat16": torch.bfloat16, |
|
"float16": torch.float16, |
|
}[dtype] |
|
ctx = ( |
|
nullcontext() |
|
if device_type == "cpu" |
|
else torch.amp.autocast(device_type=device_type, dtype=ptdtype) |
|
) |
|
|
|
|
|
if init_from == "resume": |
|
|
|
ckpt_path = os.path.normpath(f"../chess-mamba-vs-xformer/out/Mamba/{self.model_name}") |
|
checkpoint = torch.load(ckpt_path, map_location=device) |
|
model_config = checkpoint["model_args"] |
|
model = MambaLMHeadModel(model_config) |
|
model.load_state_dict(checkpoint['model']) |
|
elif init_from.startswith('state-spaces'): |
|
model = from_pretrained(init_from).to(device) |
|
else: |
|
raise ValueError("Invalid init_from value") |
|
|
|
model.eval() |
|
model.to(device) |
|
|
|
if compile and hasattr(torch, 'compile'): |
|
model = torch.compile(model) |
|
|
|
|
|
meta_path = os.path.join(BASE_DIR, "out", "meta.pkl") |
|
load_meta = os.path.exists(meta_path) |
|
if move_num_in_gamestate and load_meta: |
|
with open(meta_path, "rb") as f: |
|
meta = pickle.load(f) |
|
stoi, itos = meta["stoi"], meta["itos"] |
|
vocab_size = meta['vocab_size'] |
|
encode = lambda s: [stoi[c] for c in s] |
|
decode = lambda l: "".join([itos[i] for i in l]) |
|
else: |
|
stoi = {' ': 0, '.': 1, 'a': 2, 'b': 3, 'c': 4, 'd': 5, 'e': 6, 'f': 7, 'g': 8, 'h': 9, '1': 10, '2': 11, '3': 12, '4': 13, '5': 14, '6': 15, '7': 16, '8': 17, 'B': 18, 'N': 19, 'R': 20, 'Q': 21, 'K': 22, 'O': 23, 'x': 24, '+': 25, '#': 26, '=': 27} |
|
itos = {0: ' ', 1: '.', 2: 'a', 3: 'b', 4: 'c', 5: 'd', 6: 'e', 7: 'f', 8: 'g', 9: 'h', 10: '1', 11: '2', 12: '3', 13: '4', 14: '5', 15: '6', 16: '7', 17: '8', 18: 'B', 19: 'N', 20: 'R', 21: 'Q', 22: 'K', 23: 'O', 24: 'x', 25: '+', 26: '#', 27: '='} |
|
for s in stoi: |
|
assert itos[stoi[s]] == s |
|
vocab_size = len(stoi) |
|
print(f"Vocab size {vocab_size}") |
|
encode = lambda s: [stoi[c] for c in s.replace('-', '')] |
|
decode = lambda l: "".join([itos[i] for i in l if i < vocab_size]).replace("OOO", "O-O-O").replace("OO", "O-O") |
|
|
|
self.vocab_size = vocab_size |
|
self.encode = encode |
|
self.decode = decode |
|
self.space_tok = encode(' ')[0] |
|
self.dot_tok = encode('.')[0] |
|
self.model = model |
|
self.ctx = ctx |
|
self.device = device |
|
|
|
self.move_num = 0 |
|
self.hooks = [] |
|
self.max_seq_len = 1536 |
|
|
|
self.move_buckets = [float('inf')] |
|
|
|
if update_contrastive or update_linear: |
|
self.activations_sum = {} |
|
self.activations_count = {} |
|
if update_linear: |
|
if linear_probe_path and os.path.exists(linear_probe_path): |
|
self.linear_probes = torch.load(linear_probe_path) |
|
else: |
|
self.linear_probes = {} |
|
if update_contrastive or update_linear: |
|
linear_size = self.model.config.d_model * 8 |
|
for i, layer in enumerate(self.model.backbone.layers): |
|
self.activations_sum[i] = {bucket: {"won": np.zeros((1, 8, self.model.config.d_model)), |
|
"lost": np.zeros((1, 8, self.model.config.d_model)), |
|
"current": np.zeros((1, 8, self.model.config.d_model))} |
|
for bucket in self.move_buckets} |
|
self.activations_count[i] = {bucket: {"won": 0, "lost": 0, "current": 0} |
|
for bucket in self.move_buckets} |
|
|
|
def hook(module, input, output, layer_idx=i): |
|
if isinstance(output, tuple): |
|
tensor_output = output[0] |
|
else: |
|
tensor_output = output |
|
seq_len = tensor_output.shape[1] |
|
bucket = next(b for b in self.move_buckets if self.move_num <= b) |
|
self.activations_sum[layer_idx][bucket]["current"][:, :min(8, self.seq_len), :] += tensor_output.detach().cpu().numpy()[:, :self.seq_len, :][:, -8:, :] |
|
self.activations_count[layer_idx][bucket]["current"] += 1 |
|
|
|
self.hooks.append(layer.register_forward_hook(hook)) |
|
if update_linear: |
|
if not linear_probe_path or not os.path.exists(linear_probe_path): |
|
self.linear_probes[i] = { |
|
'q_value': nn.Linear(linear_size, 1), |
|
'q_value_delta': nn.Linear(linear_size, 1), |
|
'material_balance': nn.Linear(linear_size, 1) |
|
} |
|
if update_linear: |
|
self.linear_probe_targets = {i: {bucket: {'q_value': [], 'q_value_delta': [], 'material_balance': []} for bucket in self.move_buckets} for i in self.linear_probes} |
|
self.linear_optimizers = { |
|
layer_idx: { |
|
probe_type: optim.Adam(self.linear_probes[layer_idx][probe_type].parameters(), lr=0.01) |
|
for probe_type in ['q_value', 'q_value_delta', 'material_balance'] |
|
} |
|
for layer_idx in self.linear_probes |
|
} |
|
wandb.init(project="mamba_linear_probes", name=f"mamba_linear_probes") |
|
self.wandb_step = 0 |
|
self.linear_save_ct = 0 |
|
|
|
def get_mamba_response(self, game_state: str, temperature: float, max_new_tokens: int, top_k: int): |
|
game_state = game_state.split("\n\n")[-1].strip() |
|
|
|
|
|
|
|
encoded_prompt = self.encode(game_state) |
|
input_ids = torch.tensor([encoded_prompt], dtype=torch.long, device=self.device) |
|
self.seq_len = input_ids[0].size(dim=0) |
|
|
|
self.model.eval() |
|
with torch.no_grad(): |
|
have_non_space = False |
|
for _ in range(max_new_tokens): |
|
logits = self.model(input_ids).logits[0, -1, :] |
|
|
|
|
|
logits = logits / temperature |
|
if top_k > 0: |
|
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None] |
|
logits[indices_to_remove] = -float('Inf') |
|
|
|
probs = torch.nn.functional.softmax(logits, dim=-1) |
|
probs = torch.clamp(probs, min=1e-6, max=1.0) |
|
probs = probs / probs.sum() |
|
try: |
|
next_token_id = torch.multinomial(probs, num_samples=1) |
|
except: |
|
return None |
|
if next_token_id == self.space_tok or next_token_id==self.dot_tok: |
|
if have_non_space: |
|
break |
|
else: |
|
have_non_space = True |
|
input_ids = torch.cat([input_ids, next_token_id.unsqueeze(0)], dim=1) |
|
self.seq_len += 1 |
|
|
|
model_response = self.decode(input_ids[0].tolist()) |
|
model_response = model_response[len(game_state):].split(";")[0] |
|
return model_response |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
def get_move_from_response(self, response: str) -> str: |
|
if not response or len(response) == 0: |
|
return None |
|
|
|
try: |
|
moves = response.split() |
|
first_move = moves[0] |
|
first_move = first_move.lstrip('.') |
|
|
|
return first_move |
|
except: |
|
return None |
|
|
|
def get_move(self, board: chess.Board, game_state: str, temperature: float) -> str: |
|
self.move_num = game_state.count('.') |
|
completion = self.get_mamba_response(game_state, temperature, 8, self.vocab_size) |
|
return self.get_move_from_response(completion) |
|
|
|
def get_config(self) -> dict: |
|
return {"model": self.model_name} |
|
|
|
def update_activations(self, result): |
|
for layer_idx in self.activations_sum: |
|
if result == "reset": |
|
self.activations_sum[layer_idx] = {bucket: {"won": np.zeros((1, 8, self.model.config.d_model)), |
|
"lost": np.zeros((1, 8, self.model.config.d_model)), |
|
"current": np.zeros((1, 8, self.model.config.d_model))} |
|
for bucket in self.move_buckets} |
|
self.activations_count[layer_idx] = {bucket: {"won": 0, "lost": 0, "current": 0} |
|
for bucket in self.move_buckets} |
|
else: |
|
for bucket in self.move_buckets: |
|
self.activations_sum[layer_idx][bucket][result] += self.activations_sum[layer_idx][bucket]["current"] |
|
self.activations_count[layer_idx][bucket][result] += 1 |
|
|
|
def save_activations(self, path): |
|
if os.path.exists(path): |
|
with open(path, "rb") as f: |
|
activations_sum, activations_count = pickle.load(f) |
|
else: |
|
activations_sum = {} |
|
activations_count = {} |
|
|
|
for layer_idx in self.activations_sum: |
|
for bucket in self.move_buckets: |
|
if self.activations_count[layer_idx][bucket]["current"] == 0: |
|
continue |
|
if layer_idx not in activations_sum: |
|
activations_sum[layer_idx] = {} |
|
activations_count[layer_idx] = {} |
|
if bucket not in activations_sum[layer_idx]: |
|
activations_sum[layer_idx][bucket] = {} |
|
activations_count[layer_idx][bucket] = {} |
|
for category in ["won", "lost"]: |
|
if category not in activations_sum[layer_idx][bucket]: |
|
activations_sum[layer_idx][bucket][category] = np.zeros((1, 8, self.model.config.d_model)) |
|
activations_count[layer_idx][bucket][category] = 0 |
|
|
|
activations_sum[layer_idx][bucket][category] += self.activations_sum[layer_idx][bucket][category] |
|
activations_count[layer_idx][bucket][category] += self.activations_count[layer_idx][bucket][category] |
|
|
|
with open(path, "wb") as f: |
|
pickle.dump((activations_sum, activations_count), f) |
|
|
|
for layer_idx in self.activations_sum: |
|
self.activations_sum[layer_idx] = {bucket: {"won": np.zeros((1, 8, self.model.config.d_model)), |
|
"lost": np.zeros((1, 8, self.model.config.d_model)), |
|
"current": np.zeros((1, 8, self.model.config.d_model))} |
|
for bucket in self.move_buckets} |
|
self.activations_count[layer_idx] = {bucket: {"won": 0, "lost": 0, "current": 0} |
|
for bucket in self.move_buckets} |
|
|
|
def apply_contrastive_activations(self, path, weight=1.0): |
|
if os.path.exists(path): |
|
with open(path, "rb") as f: |
|
activations_sum, activations_count = pickle.load(f) |
|
|
|
self.contrastive_activations_cache = {} |
|
|
|
def hook(module, input, output, layer_idx): |
|
if isinstance(output, tuple): |
|
tensor_output = output[0] |
|
else: |
|
tensor_output = output |
|
seq_len = tensor_output.shape[1] |
|
bucket = next(b for b in self.move_buckets if self.move_num <= b) |
|
|
|
|
|
if layer_idx in self.contrastive_activations_cache and bucket in self.contrastive_activations_cache[layer_idx]: |
|
safe_contrastive_activations = self.contrastive_activations_cache[layer_idx][bucket] |
|
else: |
|
won_activations = activations_sum[layer_idx][bucket]["won"] / activations_count[layer_idx][bucket]["won"] |
|
lost_activations = activations_sum[layer_idx][bucket]["lost"] / activations_count[layer_idx][bucket]["lost"] |
|
contrastive_activations = won_activations - lost_activations |
|
contrastive_activations_tensor = torch.from_numpy(contrastive_activations).to(tensor_output.device) |
|
valid_activations = torch.isfinite(contrastive_activations_tensor) |
|
safe_contrastive_activations = torch.zeros_like(contrastive_activations_tensor) |
|
safe_contrastive_activations[valid_activations] = contrastive_activations_tensor[valid_activations] |
|
|
|
|
|
if layer_idx not in self.contrastive_activations_cache: |
|
self.contrastive_activations_cache[layer_idx] = {} |
|
self.contrastive_activations_cache[layer_idx][bucket] = safe_contrastive_activations |
|
|
|
tensor_output += safe_contrastive_activations[:, :seq_len, :] * weight |
|
if isinstance(output, tuple): |
|
return tensor_output, output[1] |
|
else: |
|
return tensor_output |
|
|
|
for layer_idx in activations_sum: |
|
self.hooks.append(self.model.backbone.layers[layer_idx].register_forward_hook( |
|
lambda module, input, output, layer_idx=layer_idx: hook(module, input, output, layer_idx) |
|
)) |
|
|
|
def update_linear_probe_targets(self, curr_q_value, q_value_delta, material_bal): |
|
bucket = next(b for b in self.move_buckets if self.move_num <= b) |
|
for layer_idx in self.linear_probe_targets: |
|
self.linear_probe_targets[layer_idx][bucket]['q_value'].append(curr_q_value) |
|
self.linear_probe_targets[layer_idx][bucket]['q_value_delta'].append(q_value_delta) |
|
self.linear_probe_targets[layer_idx][bucket]['material_balance'].append(material_bal) |
|
|
|
def train_linear_probes(self): |
|
def get_lr(it): |
|
warmup_iters = 0 |
|
lr_decay_iters = 3000 * 43 |
|
learning_rate = 0.0000075 |
|
min_lr = 0.00000075 |
|
|
|
if it < warmup_iters: |
|
return learning_rate * it / warmup_iters |
|
|
|
if it > lr_decay_iters: |
|
return min_lr |
|
|
|
decay_ratio = (it - warmup_iters) / (lr_decay_iters - warmup_iters) |
|
assert 0 <= decay_ratio <= 1 |
|
coeff = 0.5 * (1.0 + math.cos(math.pi * decay_ratio)) |
|
return min_lr + coeff * (learning_rate - min_lr) |
|
|
|
criterion = nn.MSELoss() |
|
self.wandb_step += 1 |
|
lr = get_lr(self.wandb_step) |
|
|
|
for layer_idx in self.linear_probes: |
|
for bucket in self.move_buckets: |
|
if self.activations_count[layer_idx][bucket]['current'] > 0: |
|
X = torch.from_numpy(self.activations_sum[layer_idx][bucket]['current']).float().flatten(1) |
|
for probe_type in ['q_value', 'q_value_delta', 'material_balance']: |
|
y = torch.tensor(self.linear_probe_targets[layer_idx][bucket][probe_type]).float().unsqueeze(1) |
|
if len(y) > 0: |
|
y_pred = self.linear_probes[layer_idx][probe_type](X) |
|
loss = criterion(y_pred, y) |
|
for param_group in self.linear_optimizers[layer_idx][probe_type].param_groups: |
|
param_group['lr'] = lr |
|
self.linear_optimizers[layer_idx][probe_type].zero_grad() |
|
loss.backward() |
|
self.linear_optimizers[layer_idx][probe_type].step() |
|
|
|
wandb.log({ |
|
"etc/lr": lr, |
|
f"{probe_type}/layer_{layer_idx}_loss": loss.item() |
|
}, step=self.wandb_step) |
|
|
|
|
|
self.linear_probe_targets = {i: {bucket: {'q_value': [], 'q_value_delta': [], 'material_balance': []} for bucket in self.move_buckets} for i in self.linear_probes} |
|
|
|
def save_linear_probe_data(self, path): |
|
self.linear_save_ct += 25 |
|
wandb.log({ |
|
"etc/games": self.linear_save_ct |
|
}, step=self.wandb_step) |
|
torch.save(self.linear_probes, path) |
|
|
|
def evaluate_linear_probes(self, board: chess.Board): |
|
self.move_num = board.fullmove_number |
|
bucket = next(b for b in self.move_buckets if self.move_num <= b) |
|
|
|
|
|
probe_stats = {probe_type: {layer_idx: {self.move_num: None} for layer_idx in self.linear_probes} for probe_type in ['q_value', 'q_value_delta', 'material_balance']} |
|
|
|
for layer_idx in self.linear_probes: |
|
X = torch.from_numpy(self.activations_sum[layer_idx][bucket]['current']).float().flatten(1) |
|
for probe_type in ['q_value', 'q_value_delta', 'material_balance']: |
|
target = torch.tensor(self.linear_probe_targets[layer_idx][bucket][probe_type]).float().item() |
|
probe = self.linear_probes[layer_idx][probe_type] |
|
prediction = probe(X).item() |
|
|
|
|
|
|
|
if probe_type == 'q_value': |
|
accuracy = 1 - abs(prediction - target) / 2 |
|
elif probe_type == 'q_value_delta': |
|
accuracy = 1 - abs(prediction - target) / 4 |
|
else: |
|
max_range = 35 |
|
accuracy = 1 - min(abs(prediction - target) / max_range, 1) |
|
|
|
|
|
probe_stats[probe_type][layer_idx][self.move_num] = accuracy |
|
|
|
self.linear_probe_targets = {i: {bucket: {'q_value': [], 'q_value_delta': [], 'material_balance': []} for bucket in self.move_buckets} for i in self.linear_probes} |
|
|
|
|
|
with open('probe_stats.json', 'a') as f: |
|
json.dump(probe_stats, f) |
|
f.write('\n') |
|
|