ce_score_job: multiple inputs, OUT_TAG

code/hf_jobs/ce_score_job.py

@@ -3,19 +3,20 @@ ce/scores/{INPUT}.parquet (s1_id, cand_id, ce_logit). Env: INPUT (e.g. score_tes
 import os, time, numpy as np, polars as pl, torch
 from huggingface_hub import hf_hub_download, snapshot_download, HfApi
 from transformers import AutoTokenizer, AutoModelForSequenceClassification
-TOK = os.environ['HF_TOKEN']; REPO = 'Arushhh/amlc-work'; INP = os.environ['INPUT']; T0 = time.time()
+TOK = os.environ['HF_TOKEN']; REPO = 'Arushhh/amlc-work'; INPUTS = os.environ['INPUT'].split(','); TAG = os.environ.get('OUT_TAG', ''); T0 = time.time()
 def log(*a): print(f'[{time.strftime("%H:%M:%S")} +{time.time()-T0:5.0f}s]', *a, flush=True)
 p = snapshot_download(os.environ.get('CE_REPO', 'Arushhh/amlc-ce'), token=TOK, local_dir='/workspace/ce')
 tok = AutoTokenizer.from_pretrained(p); model = AutoModelForSequenceClassification.from_pretrained(p).cuda().eval()
-df = pl.read_parquet(hf_hub_download(REPO, f'ce/{INP}.parquet', repo_type='dataset', token=TOK)); log('pairs', df.height)
-order = np.argsort((df['text_a'].str.len_chars() + df['text_b'].str.len_chars()).to_numpy()); A, B = df['text_a'].to_list(), df['text_b'].to_list()
-out = np.zeros(df.height, dtype=np.float32); bs = 1024; t = time.time()
-with torch.no_grad():
-    for i in range(0, len(order), bs):
-        j = order[i:i + bs]
-        enc = tok([A[k] for k in j], [B[k] for k in j], truncation='longest_first', max_length=128, padding=True, return_tensors='pt')
-        with torch.autocast('cuda', dtype=torch.bfloat16):
-            out[j] = model(**{k: v.cuda() for k, v in enc.items()}).logits.float().squeeze(-1).cpu().numpy()
-        if (i // bs) % 500 == 0: log(f'scored {i + len(j):,}/{df.height:,} ({(i + len(j)) / (time.time() - t):,.0f} pairs/s)')
-df.select('s1_id', 'cand_id').with_columns(pl.Series('ce_logit', out)).write_parquet('/workspace/out.parquet')
-HfApi(token=TOK).upload_file(path_or_fileobj='/workspace/out.parquet', path_in_repo=f'ce/scores/{INP}.parquet', repo_id=REPO, repo_type='dataset'); log('CE_SCORE_DONE', INP)
+for INP in INPUTS:
+  df = pl.read_parquet(hf_hub_download(REPO, f'ce/{INP}.parquet', repo_type='dataset', token=TOK)); log(INP, 'pairs', df.height)
+  order = np.argsort((df['text_a'].str.len_chars() + df['text_b'].str.len_chars()).to_numpy()); A, B = df['text_a'].to_list(), df['text_b'].to_list()
+  out = np.zeros(df.height, dtype=np.float32); bs = 1024; t = time.time()
+  with torch.no_grad():
+      for i in range(0, len(order), bs):
+          j = order[i:i + bs]
+          enc = tok([A[k] for k in j], [B[k] for k in j], truncation='longest_first', max_length=128, padding=True, return_tensors='pt')
+          with torch.autocast('cuda', dtype=torch.bfloat16):
+              out[j] = model(**{k: v.cuda() for k, v in enc.items()}).logits.float().squeeze(-1).cpu().numpy()
+          if (i // bs) % 500 == 0: log(f'scored {i + len(j):,}/{df.height:,} ({(i + len(j)) / (time.time() - t):,.0f} pairs/s)')
+  df.select('s1_id', 'cand_id').with_columns(pl.Series('ce_logit', out)).write_parquet('/workspace/out.parquet')
+  HfApi(token=TOK).upload_file(path_or_fileobj='/workspace/out.parquet', path_in_repo=f'ce/scores/{INP}{("__" + TAG) if TAG else ""}.parquet', repo_id=REPO, repo_type='dataset'); log('CE_SCORE_DONE', INP)