# ============================================================================== # 🏛️ CELL 11 (100% STANDALONE): Flagship MyanmarTranslator-210M-Pro # (8L Enc + 8L Dec, d=768, 210.90M Params) + Net2Wider/Deeper + Auto-Upload # ============================================================================== import os os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True" import sys, gc, math, time, json, shutil, traceback import numpy as np import sentencepiece as spm import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.checkpoint import checkpoint as grad_ckpt from torch.utils.data import Dataset, DataLoader, Sampler from safetensors.torch import save_file from huggingface_hub import HfApi, create_repo torch.backends.cuda.matmul.allow_tf32 = True torch.backends.cudnn.allow_tf32 = True device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # ── 0. Capture Exact Cell Code & Setup Incremental Logger ── try: CELL_SOURCE_CODE = get_ipython().user_ns["In"][-1] except Exception: CELL_SOURCE_CODE = "# MyanmarTranslator-210M-Pro Standalone Training Script" PRO_REPO_ID = "freococo/MyanmarTranslator-210M-Pro" WORK_DIR = "/workspace" DATA_DIR = "/workspace/nmt_dataset_memmap" TOK_DIR = "/workspace/nmt_tokenizer" TOK_FILE = os.path.join(TOK_DIR, "tokenizer.model") TEACHER_84M = "/workspace/nmt_84m_polished/best_model.pt" PRO_CKPT = "/workspace/pro_210m_checkpoints" PRO_EXPORT = "/workspace/MyanmarTranslator-210M-Pro-export" os.makedirs(PRO_CKPT, exist_ok=True) if not os.path.exists(os.path.join(PRO_CKPT, "latest_pro.pt")): with open(os.path.join(PRO_CKPT, "trainingcelloutputwitheval.txt"), "w", encoding="utf-8") as lf: lf.write("") with open(os.path.join(PRO_CKPT, "training_codes.txt"), "w", encoding="utf-8") as f: f.write(CELL_SOURCE_CODE) LOG_LINES = [] def log(msg: str = ""): print(msg, flush=True) LOG_LINES.append(str(msg)) with open(os.path.join(PRO_CKPT, "trainingcelloutputwitheval.txt"), "a", encoding="utf-8") as lf: lf.write(str(msg) + "\n") gc.collect() if torch.cuda.is_available(): torch.cuda.empty_cache() free_vram, tot_vram = torch.cuda.mem_get_info() _, _, free_disk = shutil.disk_usage(WORK_DIR) log(f"🟢 Standalone Kernel Ready for 210M-Pro | GPU Free: {free_vram/1e9:.2f}/{tot_vram/1e9:.2f} GB | Disk Free: {free_disk/1e9:.1f} GB") sp = spm.SentencePieceProcessor() sp.load(TOK_FILE) VOCAB_SIZE, UNK_ID, BOS_ID, EOS_ID, PAD_ID = 65664, 0, 1, 2, 3 TOK_2MY, TOK_2EN, TOK_DENOISE, TOK_MASK = 65536, 65537, 65538, 65539 # ── 1. Standalone Architecture Definitions ── class RMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1e-6): super().__init__() self.dim = (dim,) self.eps = eps self.weight = nn.Parameter(torch.ones(dim)) def forward(self, x: torch.Tensor) -> torch.Tensor: return F.rms_norm(x, self.dim, self.weight, self.eps) class RotaryEmbedding(nn.Module): def __init__(self, dim: int, max_seq_len: int = 1024, theta: float = 10000.0): super().__init__() inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim)) t = torch.arange(max_seq_len, dtype=torch.float32) freqs = torch.outer(t, inv_freq) self.register_buffer("cos_cached", freqs.cos()[None, None, :, :], persistent=False) self.register_buffer("sin_cached", freqs.sin()[None, None, :, :], persistent=False) def forward(self, x: torch.Tensor) -> torch.Tensor: L = x.size(2) cos = self.cos_cached[:, :, :L, :].to(dtype=x.dtype) sin = self.sin_cached[:, :, :L, :].to(dtype=x.dtype) x1, x2 = x[..., ::2], x[..., 1::2] return torch.cat([x1 * cos - x2 * sin, x1 * sin + x2 * cos], dim=-1) class SwiGLU(nn.Module): def __init__(self, d_model: int, ffn_dim: int): super().__init__() self.gate_up = nn.Linear(d_model, 2 * ffn_dim, bias=False) self.down = nn.Linear(ffn_dim, d_model, bias=False) def forward(self, x: torch.Tensor) -> torch.Tensor: gate, up = self.gate_up(x).chunk(2, dim=-1) return self.down(F.silu(gate) * up) class SelfAttention(nn.Module): def __init__(self, d_model: int, n_heads: int, is_causal: bool = False): super().__init__() self.n_heads = n_heads self.head_dim = d_model // n_heads self.is_causal = is_causal self.qkv = nn.Linear(d_model, 3 * d_model, bias=False) self.out = nn.Linear(d_model, d_model, bias=False) self.q_norm = RMSNorm(self.head_dim) self.k_norm = RMSNorm(self.head_dim) def forward(self, x: torch.Tensor, rope: RotaryEmbedding, attn_mask: torch.Tensor = None) -> torch.Tensor: B, L, D = x.shape q, k, v = self.qkv(x).chunk(3, dim=-1) q = self.q_norm(q.view(B, L, self.n_heads, self.head_dim).transpose(1, 2)) k = self.k_norm(k.view(B, L, self.n_heads, self.head_dim).transpose(1, 2)) v = v.view(B, L, self.n_heads, self.head_dim).transpose(1, 2) q, k = rope(q), rope(k) y = F.scaled_dot_product_attention(q, k, v, attn_mask=None if self.is_causal else attn_mask, is_causal=self.is_causal) return self.out(y.transpose(1, 2).contiguous().view(B, L, D)) class CrossAttention(nn.Module): def __init__(self, d_model: int, n_heads: int): super().__init__() self.n_heads = n_heads self.head_dim = d_model // n_heads self.q_proj = nn.Linear(d_model, d_model, bias=False) self.kv_proj = nn.Linear(d_model, 2 * d_model, bias=False) self.out = nn.Linear(d_model, d_model, bias=False) self.q_norm = RMSNorm(self.head_dim) self.k_norm = RMSNorm(self.head_dim) def forward(self, x: torch.Tensor, enc_out: torch.Tensor, enc_mask: torch.Tensor) -> torch.Tensor: B, T, D = x.shape S = enc_out.size(1) q = self.q_norm(self.q_proj(x).view(B, T, self.n_heads, self.head_dim).transpose(1, 2)) k, v = self.kv_proj(enc_out).chunk(2, dim=-1) k = self.k_norm(k.view(B, S, self.n_heads, self.head_dim).transpose(1, 2)) v = v.view(B, S, self.n_heads, self.head_dim).transpose(1, 2) y = F.scaled_dot_product_attention(q, k, v, attn_mask=enc_mask, is_causal=False) return self.out(y.transpose(1, 2).contiguous().view(B, T, D)) class EncoderLayer(nn.Module): def __init__(self, d_model: int, n_heads: int, ffn_dim: int): super().__init__() self.norm1 = RMSNorm(d_model) self.self_attn = SelfAttention(d_model, n_heads, is_causal=False) self.norm2 = RMSNorm(d_model) self.mlp = SwiGLU(d_model, ffn_dim) def forward(self, x: torch.Tensor, rope: RotaryEmbedding, enc_mask: torch.Tensor) -> torch.Tensor: x = x + self.self_attn(self.norm1(x), rope, enc_mask) x = x + self.mlp(self.norm2(x)) return x class DecoderLayer(nn.Module): def __init__(self, d_model: int, n_heads: int, ffn_dim: int): super().__init__() self.norm1 = RMSNorm(d_model) self.self_attn = SelfAttention(d_model, n_heads, is_causal=True) self.norm2 = RMSNorm(d_model) self.cross_attn = CrossAttention(d_model, n_heads) self.norm3 = RMSNorm(d_model) self.mlp = SwiGLU(d_model, ffn_dim) def forward(self, x: torch.Tensor, enc_out: torch.Tensor, rope: RotaryEmbedding, enc_mask: torch.Tensor) -> torch.Tensor: x = x + self.self_attn(self.norm1(x), rope, None) x = x + self.cross_attn(self.norm2(x), enc_out, enc_mask) x = x + self.mlp(self.norm3(x)) return x class MyanmarTranslator84M(nn.Module): def __init__(self, vocab_size=65664, d_model=768, n_heads=12, enc_layers=8, dec_layers=8, ffn_dim=2816, max_seq_len=1024): super().__init__() self.config = dict( vocab_size=vocab_size, d_model=d_model, n_heads=n_heads, enc_layers=enc_layers, dec_layers=dec_layers, ffn_dim=ffn_dim, max_seq_len=max_seq_len ) self.d_model = d_model self.embed_scale = math.sqrt(d_model) self.embed_tokens = nn.Embedding(vocab_size, d_model, padding_idx=PAD_ID) self.rope = RotaryEmbedding(d_model // n_heads, max_seq_len=max_seq_len) self.encoder = nn.ModuleList([EncoderLayer(d_model, n_heads, ffn_dim) for _ in range(enc_layers)]) self.enc_norm = RMSNorm(d_model) self.decoder = nn.ModuleList([DecoderLayer(d_model, n_heads, ffn_dim) for _ in range(dec_layers)]) self.dec_norm = RMSNorm(d_model) self.lm_head = nn.Linear(d_model, vocab_size, bias=False) self.lm_head.weight = self.embed_tokens.weight self._init_weights(enc_layers + dec_layers) def _init_weights(self, total_layers: int): res_std = 0.02 / math.sqrt(2 * total_layers) for name, m in self.named_modules(): if isinstance(m, nn.Linear): nn.init.normal_(m.weight, mean=0.0, std=res_std if (name.endswith(".out") or name.endswith(".down")) else 0.02) elif isinstance(m, nn.Embedding): nn.init.normal_(m.weight, mean=0.0, std=0.02) if m.padding_idx is not None: m.weight.data[m.padding_idx].zero_() def encode(self, src_ids: torch.Tensor, use_ckpt: bool = False): enc_mask = (src_ids != PAD_ID).unsqueeze(1).unsqueeze(2) x = self.embed_tokens(src_ids) * self.embed_scale for layer in self.encoder: x = grad_ckpt(layer, x, self.rope, enc_mask, use_reentrant=False) if (use_ckpt and self.training) else layer(x, self.rope, enc_mask) return self.enc_norm(x), enc_mask def decode(self, tgt_in: torch.Tensor, enc_out: torch.Tensor, enc_mask: torch.Tensor, use_ckpt: bool = False): x = self.embed_tokens(tgt_in) * self.embed_scale for layer in self.decoder: x = grad_ckpt(layer, x, enc_out, self.rope, enc_mask, use_reentrant=False) if (use_ckpt and self.training) else layer(x, enc_out, self.rope, enc_mask) return self.dec_norm(x) def forward_and_backward(self, src_ids: torch.Tensor, tgt_in: torch.Tensor, labels: torch.Tensor, loss_scale: float = 1.0, label_smoothing: float = 0.1): use_ckpt = max(src_ids.size(1), tgt_in.size(1)) > 192 with torch.amp.autocast("cuda", dtype=torch.bfloat16): enc_out, enc_mask = self.encode(src_ids, use_ckpt=use_ckpt) dec_out = self.decode(tgt_in, enc_out, enc_mask, use_ckpt=use_ckpt) valid_mask = labels.ne(-100) valid_h = dec_out[valid_mask] valid_y = labels[valid_mask] n_valid = valid_y.numel() h_det = valid_h.detach().requires_grad_(True) denom = float(n_valid) * float(loss_scale) tot_loss = 0.0 chunk_sz = 2048 for i in range(0, n_valid, chunk_sz): hc = h_det[i : i + chunk_sz] yc = valid_y[i : i + chunk_sz] with torch.amp.autocast("cuda", dtype=torch.bfloat16): logits_c = self.lm_head(hc) lc = F.cross_entropy(logits_c, yc, label_smoothing=label_smoothing, reduction="sum") (lc / denom).backward() tot_loss += float(lc.detach().item()) del logits_c, lc valid_h.backward(h_det.grad) return tot_loss / max(n_valid, 1), n_valid @torch.no_grad() def eval_nll(self, src_ids: torch.Tensor, tgt_in: torch.Tensor, labels: torch.Tensor): with torch.amp.autocast("cuda", dtype=torch.bfloat16): enc_out, enc_mask = self.encode(src_ids, use_ckpt=False) dec_out = self.decode(tgt_in, enc_out, enc_mask, use_ckpt=False) valid_mask = labels.ne(-100) valid_h = dec_out[valid_mask] valid_y = labels[valid_mask] n_valid = valid_y.numel() tot_nll = 0.0 for i in range(0, n_valid, 2048): logits_c = self.lm_head(valid_h[i : i + 2048]) tot_nll += float(F.cross_entropy(logits_c, valid_y[i : i + 2048], reduction="sum").item()) return tot_nll, n_valid @torch.no_grad() def translate(self, text: str, direction: str = "en2my", max_new_tokens: int = 192, rep_penalty: float = 1.15) -> str: self.eval() dir_tok = TOK_2MY if direction == "en2my" else TOK_2EN s_ids = [dir_tok] + sp.encode_as_ids(text.strip())[:1020] + [EOS_ID] src_t = torch.tensor([s_ids], dtype=torch.long, device=next(self.parameters()).device) with torch.amp.autocast("cuda", dtype=torch.bfloat16): enc_out, enc_mask = self.encode(src_t, use_ckpt=False) ys = [BOS_ID] for _ in range(max_new_tokens): tgt_t = torch.tensor([ys], dtype=torch.long, device=src_t.device) dec_out = self.decode(tgt_t, enc_out, enc_mask, use_ckpt=False) logits = self.lm_head(dec_out[:, -1, :]).float()[0] logits[[UNK_ID, PAD_ID, BOS_ID]] = -1e9 logits[65536:] = -1e9 if rep_penalty != 1.0 and len(ys) > 1: seen = torch.tensor(list(set(ys[1:])), dtype=torch.long, device=src_t.device) s_logits = logits[seen] logits[seen] = torch.where(s_logits > 0, s_logits / rep_penalty, s_logits * rep_penalty) nxt = int(logits.argmax().item()) if nxt == EOS_ID: break ys.append(nxt) return sp.decode_ids([x for x in ys[1:] if x < 65536]) @torch.no_grad() def translate_beam(self, text: str, direction: str = "en2my", num_beams: int = 4, max_new_tokens: int = 192, length_penalty: float = 0.85, rep_penalty: float = 1.12) -> str: self.eval() if num_beams <= 1: return self.translate(text, direction=direction, max_new_tokens=max_new_tokens, rep_penalty=rep_penalty) dir_tok = TOK_2MY if direction == "en2my" else TOK_2EN s_ids = [dir_tok] + sp.encode_as_ids(text.strip())[:1020] + [EOS_ID] src_t = torch.tensor([s_ids], dtype=torch.long, device=next(self.parameters()).device) with torch.amp.autocast("cuda", dtype=torch.bfloat16): enc_out, enc_mask = self.encode(src_t, use_ckpt=False) beams = [(0.0, [BOS_ID], False)] for _ in range(max_new_tokens): active = [(sc, seq) for sc, seq, done in beams if not done] if not active: break B_act = len(active) tgt_t = torch.tensor([seq for _, seq in active], dtype=torch.long, device=src_t.device) dec_out = self.decode(tgt_t, enc_out.expand(B_act, -1, -1), enc_mask.expand(B_act, -1, -1, -1), use_ckpt=False) logits = self.lm_head(dec_out[:, -1, :]).float() logits[:, [UNK_ID, PAD_ID, BOS_ID]] = -1e9 logits[:, 65536:] = -1e9 for b_i, (_, seq) in enumerate(active): if rep_penalty != 1.0 and len(seq) > 1: seen = torch.tensor(list(set(seq[1:])), dtype=torch.long, device=src_t.device) s_l = logits[b_i, seen] logits[b_i, seen] = torch.where(s_l > 0, s_l / rep_penalty, s_l * rep_penalty) log_probs = F.log_softmax(logits, dim=-1) topk_lp, topk_ids = torch.topk(log_probs, num_beams * 2, dim=-1) candidates = [(sc, seq, True) for sc, seq, done in beams if done] for b_i, (sc, seq) in enumerate(active): for k in range(num_beams * 2): tok_id = int(topk_ids[b_i, k].item()) new_sc = sc + float(topk_lp[b_i, k].item()) candidates.append((new_sc, seq if tok_id == EOS_ID else seq + [tok_id], tok_id == EOS_ID)) candidates.sort(key=lambda x: x[0] / (max(1, len(x[1]) - 1) ** length_penalty), reverse=True) beams = candidates[:num_beams] if all(done for _, _, done in beams): break return sp.decode_ids([x for x in beams[0][1][1:] if x < 65536]) # ── 2. Standalone Dataset & Sampler Definitions ── class Seq2SeqMemmapDataset(Dataset): def __init__(self, data_dir: str, split: str = "train"): self.src_path = os.path.join(data_dir, f"{split}_src.bin") self.tgt_path = os.path.join(data_dir, f"{split}_tgt.bin") self.index = np.load(os.path.join(data_dir, f"{split}_index.npy")) self.src_mm = None self.tgt_mm = None def _ensure_open(self): if self.src_mm is None: self.src_mm = np.memmap(self.src_path, dtype=np.uint16, mode="r") self.tgt_mm = np.memmap(self.tgt_path, dtype=np.uint16, mode="r") def __len__(self): return len(self.index) def __getitem__(self, idx: int): self._ensure_open() s0, sl, t0, tl, kind = self.index[idx] s_raw = self.src_mm[s0 : s0 + sl].astype(np.int64) if kind == 2: s_raw[s_raw == 0] = TOK_MASK src = np.empty(sl + 1, dtype=np.int64) src[0] = 65536 + kind src[1:] = s_raw tgt = self.tgt_mm[t0 : t0 + tl].astype(np.int64) return src, tgt class LengthBucketBatchSampler(Sampler): def __init__(self, index_arr: np.ndarray, batch_size: int, mega_mult: int = 256, shuffle: bool = True, seed: int = 42, start_batch: int = 0): self.lengths = (index_arr[:, 1] + index_arr[:, 3]).astype(np.int32) self.batch_size = batch_size self.mega_size = batch_size * mega_mult self.shuffle = shuffle self.seed = seed self.start_batch = start_batch self.num_batches = len(self.lengths) // batch_size def __len__(self): return max(0, self.num_batches - self.start_batch) def __iter__(self): rng = np.random.RandomState(self.seed) n = len(self.lengths) order = rng.permutation(n) if self.shuffle else np.arange(n) batches = [] for i in range(0, n - self.batch_size + 1, self.mega_size): chunk = order[i : i + self.mega_size] chunk = chunk[np.argsort(self.lengths[chunk], kind="mergesort")] for j in range(0, len(chunk) - self.batch_size + 1, self.batch_size): batches.append(chunk[j : j + self.batch_size]) if self.shuffle: rng.shuffle(batches) for b in batches[self.start_batch:]: yield b.tolist() def collate_nmt(batch): B = len(batch) max_s = min(1024, ((max(len(x[0]) for x in batch) + 15) // 16) * 16) max_t = min(1024, ((max(len(x[1]) - 1 for x in batch) + 15) // 16) * 16) src_ids = torch.full((B, max_s), PAD_ID, dtype=torch.long) tgt_in = torch.full((B, max_t), PAD_ID, dtype=torch.long) labels = torch.full((B, max_t), -100, dtype=torch.long) for i, (s, t) in enumerate(batch): sl = min(len(s), max_s) tl = min(len(t) - 1, max_t) src_ids[i, :sl] = torch.from_numpy(s[:sl]) tgt_in[i, :tl] = torch.from_numpy(t[:tl]) labels[i, :tl] = torch.from_numpy(t[1 : tl + 1]) return src_ids, tgt_in, labels AUDIT_SAMPLES = [ ("En->My Formal", "en2my", "The central bank raised interest rates to control inflation and stabilize the economy."), ("En->My Spoken", "en2my", "Honestly, I didn't expect things to turn out like this, but don't worry—we'll figure it out together."), ("My->En Formal", "my2en", "ဗဟိုဘဏ်သည် ငွေကြေးဖောင်းပွမှုကို ထိန်းချုပ်ရန်နှင့် စီးပွားရေးတည်ငြိမ်စေရန် အတိုးနှုန်းများကို မြှင့်တင်ခဲ့သည်။"), ("My->En Spoken", "my2en", "အချစ်ရေ မင်းနဲ့ ငါနဲ့ သိပ်ချစ်ခဲ့တာ ဘုရားလည်း သိတယ်၊ သကြားလည်း သိတယ်၊ နတ်လည်း သိတယ်၊ မနာလိုသူတွေ ဂျင်းထည့်တာ ယုံနေတော့ မင်း ရူးသွားပြီပဲ။"), ] UNSEEN_TESTS = [ ("En->My News", "en2my", "Heavy monsoon rains caused severe flooding in several townships, forcing thousands of residents to evacuate to temporary shelters."), ("En->My Health", "en2my", "Doctors recommend drinking plenty of clean water and getting enough rest to recover quickly from seasonal flu."), ("En->My Spoken", "en2my", "I was stuck in traffic for over an hour this morning, so please start the meeting without me."), ("My->En News", "my2en", "နိုင်ငံတကာ ကုန်သွယ်ရေး တိုးတက်လာစေရန်အတွက် နယ်စပ်ကုန်သွယ်ရေးစခန်းများတွင် ကုန်ပစ္စည်းစစ်ဆေးမှုများကို ပိုမိုမြန်ဆန်အောင် ဆောင်ရွက်နေပါသည်။"), ("My->En Tech", "my2en", "ဉာဏ်ရည်တုနည်းပညာသည် ပညာရေး၊ ကျန်းမာရေးနှင့် စီးပွားရေးလုပ်ငန်းများတွင် လုပ်ငန်းစဉ်များကို ပိုမိုထိရောက်စေရန် ကူညီပေးနိုင်သည်။"), ("My->En Spoken", "my2en", "မနက်ဖြန် မနက်စောစော သွားစရာရှိလို့ ဒီညတော့ စောစောအိပ်တော့မယ်နော်၊ မနက်ကျမှ ဖုန်းဆက်လိုက်မယ်။"), ] SPOKEN_IDIOM_PAIRS_PRO = [ ("my2en", "အချစ်ရေ မင်းနဲ့ ငါနဲ့ သိပ်ချစ်ခဲ့တာ ဘုရားလည်း သိတယ်၊ သကြားလည်း သိတယ်၊ နတ်လည်း သိတယ်၊ မနာလိုသူတွေ ဂျင်းထည့်တာ ယုံနေတော့ မင်း ရူးသွားပြီပဲ။", "My love, heaven and earth know how deeply you and I loved each other, yet because you believed the lies fed by jealous people, you've lost your mind."), ("my2en", "အချစ်ရေ မင်းနဲ့ ငါနဲ့ သိပ်ချစ်ခဲ့တာ ဘုရားလည်း သိတယ်၊ သိကြားလည်း သိတယ်၊ နတ်လည်း သိတယ်၊ မနာလိုသူတွေ ဂျင်းထည့်တာ ယုံနေတော့ မင်း ယူသွားခဲ့တာပဲ။", "My love, Buddha, Sakra, and all the spirits know how deeply you and I loved each other, but because you believed the lies stirred up by jealous people, you ended up marrying someone else."), ("my2en", "သူများတွေ ဂျင်းထည့်တာကို လွယ်လွယ်မယုံပါနဲ့၊ ကိုယ်တိုင်သေချာစုံစမ်းပြီးမှ ဆုံးဖြတ်ပါ။", "Don't easily believe it when people try to trick you or feed you lies; look into it carefully yourself before deciding."), ("my2en", "ငါ မင်းအပေါ် ဘယ်လောက်သစ္စာရှိခဲ့လဲဆိုတာ ဘုရားသိ သိကြားသိပါကွာ။", "Heaven and earth know just how loyal I have been to you."), ("my2en", "မနာလိုတဲ့လူတွေ ဘေးကနေ သွေးထိုးပြီး ဂျင်းထည့်နေတာကို မင်း ဘာလို့ နားယောင်နေရတာလဲ။", "Why are you falling for the lies and manipulation of jealous people trying to stir up trouble from the sidelines?"), ("my2en", "စိတ်မပူပါနဲ့ သူငယ်ချင်းရာ၊ အခက်အခဲဆိုတာ ခဏပဲ၊ ငါတို့ အတူတူ ရင်ဆိုင်ကြတာပေါ့။", "Don't worry, my friend—hardships are only temporary, and we'll face this together."), ("my2en", "နေကောင်းရဲ့လား မိတ်ဆွေ၊ မတွေ့ရတာတောင် တော်တော်ကြာပြီနော်။", "How have you been doing, my friend? It's been quite a long time since we last met."), ("en2my", "Honestly, I didn't expect things to turn out like this, but don't worry—we'll figure it out together.", "ရိုးရိုးသားသားပြောရရင် ဒီလိုဖြစ်လာမယ်လို့ မမျှော်လင့်ထားခဲ့ပေမယ့် စိတ်မပူပါနဲ့၊ ကျွန်တော်တို့ အတူတူ အဖြေရှာကြတာပေါ့။"), ("en2my", "Don't let jealous people trick you or feed you lies; trust what we've built together.", "မနာလိုသူတွေ ဂျင်းထည့်တာကို မယုံမိပါစေနဲ့၊ ငါတို့နှစ်ယောက် အတူတူ တည်ဆောက်ခဲ့တဲ့ အရာကိုပဲ ယုံကြည်ပေးပါ။"), ("en2my", "The central bank raised interest rates to control inflation and stabilize the economy.", "ဗဟိုဘဏ်သည် ငွေကြေးဖောင်းပွမှုကို ထိန်းချုပ်ရန်နှင့် စီးပွားရေးတည်ငြိမ်စေရန် အတိုးနှုန်းများကို မြှင့်တင်ခဲ့သည်။"), ("my2en", "ဗဟိုဘဏ်သည် ငွေကြေးဖောင်းပွမှုကို ထိန်းချုပ်ရန်နှင့် စီးပွားရေးတည်ငြိမ်စေရန် အတိုးနှုန်းများကို မြှင့်တင်ခဲ့သည်။", "The central bank raised interest rates to control inflation and stabilize the economy."), ("en2my", "I was stuck in traffic for over an hour this morning, so please start the meeting without me.", "ဒီမနက် တစ်နာရီကျော်လောက် ကားပိတ်မိနေလို့ ကျွန်တော်မပါဘဲ အစည်းအဝေးကို စထားလိုက်ပါနော်။"), ("en2my", "Please go ahead and have dinner without me; I'll be home a bit late tonight.", "ဒီည အိမ်ပြန်နည်းနည်းနောက်ကျမှာမို့လို့ ကျွန်တော်မပါဘဲ ညစာ အရင်စားနှင့်ကြပါနော်။"), ] # ── 3. Initialize 210M-Pro & Run All-Layer Net2Wider+Net2Deeper Transfer ── pro_model = MyanmarTranslator84M( vocab_size=VOCAB_SIZE, d_model=768, n_heads=12, enc_layers=8, dec_layers=8, ffn_dim=2816, max_seq_len=1024 ).to(device) n_pro_params = sum(p.numel() for p in pro_model.parameters()) log(f"🏛️ MyanmarTranslator-210M-Pro Initialized: {n_pro_params:,} parameters ({n_pro_params/1e6:.2f}M) | 8L Encoder + 8L Decoder (d=768)") def transfer_84m_to_210m(student: nn.Module, teacher_ckpt_path: str): if not os.path.exists(teacher_ckpt_path): log("⚠️ Teacher checkpoint not found, using standard initialization.") return t_sd = torch.load(teacher_ckpt_path, map_location=device)["model"] scale_d = math.sqrt(512.0 / 768.0) with torch.no_grad(): student.embed_tokens.weight.data.normal_(0.0, 0.004) student.embed_tokens.weight.data[:, :512].copy_(t_sd["embed_tokens.weight"] * scale_d) student.embed_tokens.weight.data[PAD_ID].zero_() def copy_rmsnorm(s_norm, prefix): s_norm.weight.data.fill_(1.0) s_norm.weight.data[:512].copy_(t_sd[f"{prefix}.weight"] * scale_d) def copy_self_attn(s_attn, prefix): s_attn.qkv.weight.data.normal_(0.0, 0.008) s_attn.out.weight.data.normal_(0.0, 0.002) tq, tk, tv = t_sd[f"{prefix}.qkv.weight"].chunk(3, dim=0) sq, sk, sv = s_attn.qkv.weight.data.chunk(3, dim=0) sq[:512, :512].copy_(tq) sk[:512, :512].copy_(tk) sv[:512, :512].copy_(tv) s_attn.out.weight.data[:512, :512].copy_(t_sd[f"{prefix}.out.weight"]) s_attn.q_norm.weight.data.copy_(t_sd[f"{prefix}.q_norm.weight"]) s_attn.k_norm.weight.data.copy_(t_sd[f"{prefix}.k_norm.weight"]) def copy_cross_attn(s_attn, prefix): s_attn.q_proj.weight.data.normal_(0.0, 0.008) s_attn.kv_proj.weight.data.normal_(0.0, 0.008) s_attn.out.weight.data.normal_(0.0, 0.002) s_attn.q_proj.weight.data[:512, :512].copy_(t_sd[f"{prefix}.q_proj.weight"]) tk, tv = t_sd[f"{prefix}.kv_proj.weight"].chunk(2, dim=0) sk, sv = s_attn.kv_proj.weight.data.chunk(2, dim=0) sk[:512, :512].copy_(tk) sv[:512, :512].copy_(tv) s_attn.out.weight.data[:512, :512].copy_(t_sd[f"{prefix}.out.weight"]) s_attn.q_norm.weight.data.copy_(t_sd[f"{prefix}.q_norm.weight"]) s_attn.k_norm.weight.data.copy_(t_sd[f"{prefix}.k_norm.weight"]) def copy_swiglu(s_mlp, prefix): s_mlp.gate_up.weight.data.normal_(0.0, 0.008) s_mlp.down.weight.data.normal_(0.0, 0.002) tg, tu = t_sd[f"{prefix}.gate_up.weight"].chunk(2, dim=0) sg, su = s_mlp.gate_up.weight.data.chunk(2, dim=0) sg[:1728, :512].copy_(tg) su[:1728, :512].copy_(tu) s_mlp.down.weight.data[:512, :1728].copy_(t_sd[f"{prefix}.down.weight"]) for i in range(6): copy_rmsnorm(student.encoder[i].norm1, f"encoder.{i}.norm1") copy_self_attn(student.encoder[i].self_attn, f"encoder.{i}.self_attn") copy_rmsnorm(student.encoder[i].norm2, f"encoder.{i}.norm2") copy_swiglu(student.encoder[i].mlp, f"encoder.{i}.mlp") for i in range(6, 8): student.encoder[i].self_attn.out.weight.data.normal_(0.0, 0.001) student.encoder[i].mlp.down.weight.data.normal_(0.0, 0.001) copy_rmsnorm(student.enc_norm, "enc_norm") for i in range(6): copy_rmsnorm(student.decoder[i].norm1, f"decoder.{i}.norm1") copy_self_attn(student.decoder[i].self_attn, f"decoder.{i}.self_attn") copy_rmsnorm(student.decoder[i].norm2, f"decoder.{i}.norm2") copy_cross_attn(student.decoder[i].cross_attn, f"decoder.{i}.cross_attn") copy_rmsnorm(student.decoder[i].norm3, f"decoder.{i}.norm3") copy_swiglu(student.decoder[i].mlp, f"decoder.{i}.mlp") for i in range(6, 8): student.decoder[i].self_attn.out.weight.data.normal_(0.0, 0.001) student.decoder[i].cross_attn.out.weight.data.normal_(0.0, 0.001) student.decoder[i].mlp.down.weight.data.normal_(0.0, 0.001) copy_rmsnorm(student.dec_norm, "dec_norm") del t_sd gc.collect() torch.cuda.empty_cache() log("✅ Full All-Layer Net2Wider + Net2Deeper Transfer (84M -> 210M) Complete!") transfer_84m_to_210m(pro_model, TEACHER_84M) # ── 4. DataLoader, Optimizer & Schedule (24,000 Expansion Steps + 1,000 Cooldown Steps) ── MICRO_BS_PRO = 128 GRAD_ACCUM_PRO = 2 # Global BS = 256 sequences/step (~80.6k tokens/step) MAX_TOK_CHUNK_PRO = 16384 # Caps peak VRAM at ~17-19 GB on RTX 5090 (zero OOM risk!) train_ds_pro = Seq2SeqMemmapDataset(DATA_DIR, "train") val_ds_pro = Seq2SeqMemmapDataset(DATA_DIR, "val") STAGE1_STEPS_PRO = 24000 # ~2.0B expansion tokens starting from Val PPL 6.73 (= 58,000 effective steps!) WARMUP_STEPS_PRO = 1000 PEAK_LR_PRO = 3.0e-4 MIN_LR_PRO = 3.0e-5 EVAL_EVERY_PRO = 2000 LOG_EVERY_PRO = 200 decay_pro = [p for n, p in pro_model.named_parameters() if p.requires_grad and p.ndim >= 2] nodecay_pro = [p for n, p in pro_model.named_parameters() if p.requires_grad and p.ndim < 2] pro_opt = torch.optim.AdamW( [{"params": decay_pro, "weight_decay": 0.01}, {"params": nodecay_pro, "weight_decay": 0.0}], lr=PEAK_LR_PRO, betas=(0.9, 0.98), eps=1e-8, fused=True ) def get_pro_lr(step: int) -> float: if step < WARMUP_STEPS_PRO: return PEAK_LR_PRO * (step + 1) / WARMUP_STEPS_PRO prog = min(1.0, (step - WARMUP_STEPS_PRO) / max(1, STAGE1_STEPS_PRO - WARMUP_STEPS_PRO)) return MIN_LR_PRO + 0.5 * (PEAK_LR_PRO - MIN_LR_PRO) * (1.0 + math.cos(math.pi * prog)) start_step_pro = 0 best_pro_nll = float("inf") latest_pro_ckpt = os.path.join(PRO_CKPT, "latest_pro.pt") if os.path.exists(latest_pro_ckpt): ckpt_p = torch.load(latest_pro_ckpt, map_location=device) pro_model.load_state_dict(ckpt_p["model"]) pro_opt.load_state_dict(ckpt_p["optimizer"]) start_step_pro = int(ckpt_p["step"]) best_pro_nll = float(ckpt_p.get("best_val_nll", float("inf"))) log(f"🔄 Resumed 210M-Pro from step {start_step_pro:,} (best_val_nll={best_pro_nll:.4f})") train_sampler_pro = LengthBucketBatchSampler( train_ds_pro.index, batch_size=MICRO_BS_PRO, shuffle=True, seed=2026, start_batch=start_step_pro * GRAD_ACCUM_PRO ) val_sampler_pro = LengthBucketBatchSampler( val_ds_pro.index, batch_size=MICRO_BS_PRO, shuffle=True, seed=12345, start_batch=0 ) train_loader_pro = DataLoader(train_ds_pro, batch_sampler=train_sampler_pro, collate_fn=collate_nmt, num_workers=6, pin_memory=True, persistent_workers=True) val_loader_pro = DataLoader(val_ds_pro, batch_sampler=val_sampler_pro, collate_fn=collate_nmt, num_workers=2, pin_memory=True) @torch.no_grad() def evaluate_pro(max_batches: int = 80): pro_model.eval() tot_nll, tot_tok = 0.0, 0 for i, (src_ids, tgt_in, labels) in enumerate(val_loader_pro): if i >= max_batches: break B, S = src_ids.shape T = tgt_in.size(1) sub_bs = max(4, min(B, MAX_TOK_CHUNK_PRO // max(S, T, 1))) for s_i in range(0, B, sub_bs): sc = src_ids[s_i : s_i + sub_bs].to(device, non_blocking=True) tc = tgt_in[s_i : s_i + sub_bs].to(device, non_blocking=True) yc = labels[s_i : s_i + sub_bs].to(device, non_blocking=True) nll_s, n_v = pro_model.eval_nll(sc, tc, yc) tot_nll += nll_s tot_tok += n_v avg_nll = tot_nll / max(tot_tok, 1) return avg_nll, math.exp(min(avg_nll, 20.0)) if start_step_pro == 0: init_pro_nll, init_pro_ppl = evaluate_pro() best_pro_nll = init_pro_nll log(f"🔎 Step 0 (Post-Net2Wider/Deeper Warm-Start) | Val NLL: {init_pro_nll:.4f} | Val PPL: {init_pro_ppl:.2f}") encoded_idioms_pro = [] for d_code, s_txt, t_txt in SPOKEN_IDIOM_PAIRS_PRO: dir_tok = TOK_2MY if d_code == "en2my" else TOK_2EN s_arr = np.asarray([dir_tok] + sp.encode_as_ids(s_txt.strip()) + [EOS_ID], dtype=np.int64) t_arr = np.asarray([BOS_ID] + sp.encode_as_ids(t_txt.strip()) + [EOS_ID], dtype=np.int64) encoded_idioms_pro.append((s_arr, t_arr)) # ── 5. Stage 1 Expansion Training Loop (24,000 Steps) ── log(f"\n>>> Starting Flagship MyanmarTranslator-210M-Pro Stage 1 (Step {start_step_pro:,} -> {STAGE1_STEPS_PRO:,})...") pro_model.train() pro_opt.zero_grad(set_to_none=True) step = start_step_pro micro_idx = start_step_pro * GRAD_ACCUM_PRO run_loss = run_toks = 0.0 t_start = time.time() for src_ids, tgt_in, labels in train_loader_pro: micro_idx += 1 B, S = src_ids.shape T = tgt_in.size(1) max_len = max(S, T) sub_bs = max(4, min(B, MAX_TOK_CHUNK_PRO // max(max_len, 1))) n_subs = math.ceil(B / sub_bs) for s_i in range(0, B, sub_bs): s_chunk = src_ids[s_i : s_i + sub_bs].to(device, non_blocking=True) t_chunk = tgt_in[s_i : s_i + sub_bs].to(device, non_blocking=True) y_chunk = labels[s_i : s_i + sub_bs].to(device, non_blocking=True) loss_val, n_val = pro_model.forward_and_backward( s_chunk, t_chunk, y_chunk, loss_scale=(GRAD_ACCUM_PRO * n_subs), label_smoothing=0.1 ) run_loss += loss_val / n_subs run_toks += n_val + int((s_chunk != PAD_ID).sum().item()) if micro_idx % GRAD_ACCUM_PRO == 0: lr = get_pro_lr(step) for pg in pro_opt.param_groups: pg["lr"] = lr grad_norm = torch.nn.utils.clip_grad_norm_(pro_model.parameters(), 1.0) pro_opt.step() pro_opt.zero_grad(set_to_none=True) step += 1 if step % LOG_EVERY_PRO == 0: elapsed = time.time() - t_start avg_l = run_loss / (LOG_EVERY_PRO * GRAD_ACCUM_PRO) est_nll = max(0.0, (avg_l - 0.1 * math.log(VOCAB_SIZE)) / 0.9) tok_s = run_toks / max(elapsed, 1e-3) peak_gb = torch.cuda.max_memory_allocated() / 1e9 log(f"Step {step:6d}/{STAGE1_STEPS_PRO} | SmoothLoss: {avg_l:.4f} | EstNLL: {est_nll:.4f} (PPL {math.exp(min(est_nll,15)):6.2f}) | LR: {lr:.2e} | Grad: {grad_norm:.2f} | {tok_s/1e3:.1f}k tok/s | VRAM: {peak_gb:.1f}G") run_loss = run_toks = 0.0 t_start = time.time() if step % EVAL_EVERY_PRO == 0 or step == STAGE1_STEPS_PRO: val_nll, val_ppl = evaluate_pro() log("\n" + "-"*85) log(f"🔎 210M-PRO EVAL @ Step {step:,} | Val NLL: {val_nll:.4f} | Val PPL: {val_ppl:.2f}") for tag, d_code, prompt_txt in AUDIT_SAMPLES: pred_txt = pro_model.translate(prompt_txt, d_code) log(f" [{tag}]: {pred_txt}") torch.save({"step": step, "model": pro_model.state_dict(), "optimizer": pro_opt.state_dict(), "best_val_nll": best_pro_nll, "config": pro_model.config}, latest_pro_ckpt) if val_nll < best_pro_nll: best_pro_nll = val_nll torch.save({"step": step, "model": pro_model.state_dict(), "val_nll": val_nll, "config": pro_model.config}, os.path.join(PRO_CKPT, "best_pro.pt")) log(f" 🌟 New Best 210M-Pro Checkpoint Saved! (Val NLL: {best_pro_nll:.4f})") log("-"*85 + "\n") pro_model.train() t_start = time.time() if step >= STAGE1_STEPS_PRO: break # ── 6. Stage 2 Unsmoothed Cooldown & Idiom Polish (1,000 Steps, label_smoothing=0.0) ── log(f"\n>>> Starting 210M-Pro Stage 2 Cooldown (1,000 steps | label_smoothing=0.0 | LR 3e-5 -> 4e-6)...") best_pro_s1 = torch.load(os.path.join(PRO_CKPT, "best_pro.pt"), map_location=device) pro_model.load_state_dict(best_pro_s1["model"]) pro_s2_opt = torch.optim.AdamW( [{"params": decay_pro, "weight_decay": 0.01}, {"params": nodecay_pro, "weight_decay": 0.0}], lr=3e-5, betas=(0.9, 0.98), eps=1e-8, fused=True ) clean_mask_pro = (train_ds_pro.index[:, 4] < 2) & (train_ds_pro.index[:, 1] >= 16) & (train_ds_pro.index[:, 1] <= 256) clean_idx_pro = np.flatnonzero(clean_mask_pro) rng_pro = np.random.RandomState(2026) pro_model.train() for s2_step in range(1, 1001): prog = s2_step / 1000.0 lr = 4e-6 + 0.5 * (3e-5 - 4e-6) * (1.0 + math.cos(math.pi * prog)) for pg in pro_s2_opt.param_groups: pg["lr"] = lr picked = rng_pro.choice(clean_idx_pro, size=88, replace=False) batch_items = [train_ds_pro[int(ix)] for ix in picked] for _ in range(8): batch_items.append(encoded_idioms_pro[int(rng_pro.randint(len(encoded_idioms_pro)))]) src_ids, tgt_in, labels = collate_nmt(batch_items) pro_s2_opt.zero_grad(set_to_none=True) pro_model.forward_and_backward( src_ids.to(device, non_blocking=True), tgt_in.to(device, non_blocking=True), labels.to(device, non_blocking=True), loss_scale=1.0, label_smoothing=0.0 ) torch.nn.utils.clip_grad_norm_(pro_model.parameters(), 1.0) pro_s2_opt.step() if s2_step % 250 == 0: val_nll, val_ppl = evaluate_pro() log(f"🔎 210M-Pro Stage 2 Step {s2_step}/1000 | Val NLL: {val_nll:.4f} (PPL {val_ppl:.2f})") if val_nll <= best_pro_nll + 0.01: best_pro_nll = min(best_pro_nll, val_nll) torch.save({"step": STAGE1_STEPS_PRO + s2_step, "model": pro_model.state_dict(), "val_nll": val_nll, "config": pro_model.config}, os.path.join(PRO_CKPT, "best_pro.pt")) pro_model.train() # ── 7. Final Benchmark & Private HF Upload: freococo/MyanmarTranslator-210M-Pro ── best_pro_final = torch.load(os.path.join(PRO_CKPT, "best_pro.pt"), map_location=device) pro_model.load_state_dict(best_pro_final["model"]) pro_model.eval() final_pro_nll = float(best_pro_final["val_nll"]) final_pro_ppl = math.exp(final_pro_nll) log("\n" + "="*90) log(f"🏆 MYANMARTRANSLATOR-210M-PRO FINAL BENCHMARK | Val NLL: {final_pro_nll:.4f} (PPL {final_pro_ppl:.2f})") log("="*90) for tag, d_code, prompt_txt in AUDIT_SAMPLES + UNSEEN_TESTS: pred_txt = pro_model.translate_beam(prompt_txt, d_code, num_beams=4) log(f" [{tag} (Beam=4)]: {pred_txt}") log("="*90) CKPT_SUB_PRO = os.path.join(PRO_EXPORT, "checkpoint-step-25000") shutil.rmtree(PRO_EXPORT, ignore_errors=True) os.makedirs(CKPT_SUB_PRO, exist_ok=True) sd_fp32 = {k: v.detach().cpu().float().contiguous() for k, v in pro_model.state_dict().items() if k != "lm_head.weight"} sd_bf16 = {k: v.to(torch.bfloat16).contiguous() for k, v in sd_fp32.items()} save_file(sd_bf16, os.path.join(PRO_EXPORT, "model.safetensors")) save_file(sd_fp32, os.path.join(CKPT_SUB_PRO, "model.safetensors")) if os.path.exists(latest_pro_ckpt): os.remove(latest_pro_ckpt) torch.save(best_pro_final, os.path.join(CKPT_SUB_PRO, "training_state.pt")) BASE_EXPORT = "/workspace/MyanmarTranslator-84M-Base-export" for folder, dt in [(PRO_EXPORT, "bfloat16"), (CKPT_SUB_PRO, "float32")]: for fn in ["configuration_myanmartranslator.py", "modeling_myanmartranslator.py", "generation_config.json", "tokenizer.model", "tokenizer_config.json", "special_tokens_map.json"]: shutil.copyfile(os.path.join(BASE_EXPORT, fn), os.path.join(folder, fn)) with open(os.path.join(BASE_EXPORT, "config.json"), "r") as f: cfg_d = json.load(f) cfg_d.update(pro_model.config) cfg_d["torch_dtype"] = dt cfg_d["val_nll"] = round(final_pro_nll, 4) cfg_d["val_ppl"] = round(final_pro_ppl, 2) with open(os.path.join(folder, "config.json"), "w", encoding="utf-8") as f: json.dump(cfg_d, f, indent=2) full_output_log = "\n".join(LOG_LINES) + "\n" for folder in (PRO_EXPORT, CKPT_SUB_PRO): with open(os.path.join(folder, "training_codes.txt"), "w", encoding="utf-8") as f: f.write(CELL_SOURCE_CODE) with open(os.path.join(folder, "trainingcelloutputwitheval.txt"), "w", encoding="utf-8") as f: f.write(full_output_log) with open(os.path.join(PRO_EXPORT, "README.md"), "w", encoding="utf-8") as f: f.write(f"""--- language: - my - en license: apache-2.0 library_name: transformers tags: - Translation - Safetensors - Burmese - English - MyanmarLLM - flagship - custom_code --- # MyanmarTranslator-210M-Pro Flagship `210.90M`-parameter **8-Layer Bidirectional Encoder + 8-Layer Cross-Attentive Decoder** (`d_model=768`, `n_heads=12`, `ffn_dim=2816` SwiGLU) for **English ↔ Burmese (`my`)**, warm-started via All-Layer Net2Wider+Net2Deeper expansion from `MyanmarTranslator-84M-Base`. - **Validation NLL:** `{final_pro_nll:.4f}` (**`Val PPL: {final_pro_ppl:.2f}`**) - **Reproducibility Artifacts:** `training_codes.txt` & `trainingcelloutputwitheval.txt` """) api = HfApi() create_repo(PRO_REPO_ID, repo_type="model", private=True, exist_ok=True) api.upload_folder( folder_path=PRO_EXPORT, repo_id=PRO_REPO_ID, repo_type="model", commit_message=f"Release Flagship MyanmarTranslator-210M-Pro (Val PPL {final_pro_ppl:.2f} + code & logs)", ) log(f"\n🔒🚀 SUCCESS! Flagship MyanmarTranslator-210M-Pro Live at: https://huggingface.co/{PRO_REPO_ID}")