cybertronak commited on
Commit
7c44470
·
verified ·
1 Parent(s): fd99bf9

eval results 20/20

Browse files
code/ck_discover.py CHANGED
@@ -3,7 +3,7 @@ import sys, json, re, gc, math, torch
3
  import circuitkit as ck
4
  cfg = json.loads(sys.argv[1]); out = sys.argv[2]
5
  print("visible gpus:", torch.cuda.device_count(), "| config:", cfg, flush=True)
6
- model = ck.load_model("CohereLabs/tiny-aya-global", dtype="float16", algorithm=cfg["algo"])
7
  tk = model.tokenizer; MARK = "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>"
8
  try: # Tiny Aya's default chat template prepends a long system preamble to every prompt; the task would carry it into every position of the attribution buffers
9
  probe = tk.apply_chat_template([{"role": "user", "content": "XQXQ"}], tokenize=False, add_generation_prompt=True)
@@ -24,7 +24,8 @@ try: # TransformerLen
24
  else: print("W_U differs from W_E.T, left untouched", flush=True)
25
  except Exception as e: print("tie skipped:", repr(e)[:200], flush=True)
26
  gc.collect(); torch.cuda.empty_cache()
27
- print("GPU memory after load: allocated GB", round(torch.cuda.memory_allocated() / 1e9, 2), flush=True)
 
28
  try: torch.cuda.memory._record_memory_history(max_entries=200000)
29
  except Exception as e: print("memory history unavailable:", repr(e)[:100], flush=True)
30
  def _who_holds(): # on failure: biggest live allocations grouped by the code that made them
 
3
  import circuitkit as ck
4
  cfg = json.loads(sys.argv[1]); out = sys.argv[2]
5
  print("visible gpus:", torch.cuda.device_count(), "| config:", cfg, flush=True)
6
+ model = ck.load_model("CohereLabs/tiny-aya-global", dtype=cfg.get("dtype", "float16"), algorithm=cfg["algo"], **({"device": cfg["device"]} if cfg.get("device") else {}))
7
  tk = model.tokenizer; MARK = "<|START_OF_TURN_TOKEN|><|USER_TOKEN|>"
8
  try: # Tiny Aya's default chat template prepends a long system preamble to every prompt; the task would carry it into every position of the attribution buffers
9
  probe = tk.apply_chat_template([{"role": "user", "content": "XQXQ"}], tokenize=False, add_generation_prompt=True)
 
24
  else: print("W_U differs from W_E.T, left untouched", flush=True)
25
  except Exception as e: print("tie skipped:", repr(e)[:200], flush=True)
26
  gc.collect(); torch.cuda.empty_cache()
27
+ try: print("GPU memory after load: allocated GB", round(torch.cuda.memory_allocated() / 1e9, 2), flush=True)
28
+ except Exception: pass
29
  try: torch.cuda.memory._record_memory_history(max_entries=200000)
30
  except Exception as e: print("memory history unavailable:", repr(e)[:100], flush=True)
31
  def _who_holds(): # on failure: biggest live allocations grouped by the code that made them
results/S_G_gradattr_FP16_W4.json ADDED
The diff for this file is too large to render. See raw diff
 
results/S_G_gradattr_W8_W4.json ADDED
The diff for this file is too large to render. See raw diff