runs again

_safetensors.py

@@ -1,20 +1,31 @@
 # ran into memory issues with safetensors. this code moves by them.
-import contextlib, os
+import contextlib, json, mmap, os
+
+import torch
+
+# huggingface has a size limit on individual files
+# this should shard its files
+# the filename would be some kind of base name that would turn into the index name
+# presently state is tracked assuming one file
+# an organized approach would be to separate it into two classes
+
+# what would the internal interface provided to the outer class look like?
+
 
 class WritingSafeTensors:
-    def __init__(self, filename, **metadata):
+    def __init__(self, filename, max_size=16*1024*1024*1024, **metadata):
         self.filename = filename
         self.fd = os.open(self.filename, os.O_RDWR | os.O_CREAT)
         self.size = 0
         self.capacity = 0
         self.mmapview = None
-        self.header = {'__metadata__': metadata}
+        self.header = {'__metadata__': {k:str(v) for k,v in metadata.items()}}
     def _reserve(self, length):
         if self.size + length > self.capacity:
             new_capacity = self.size * 2
             if new_capacity < self.size + length:
                 new_capacity = (((self.size + length)*2 - 1) // mmap.PAGESIZE + 1) * mmap.PAGESIZE
-                os.truncate(self.fn, new_capacity)
+            os.truncate(self.filename, new_capacity)
             self.mmapview = memoryview(mmap.mmap(self.fd, new_capacity))
             self.capacity = new_capacity
     def add(self, name, tensor):
@@ -23,10 +34,10 @@ class WritingSafeTensors:
         self._reserve(length)
         torch.frombuffer(
                 self.mmapview[self.size:self.size+length],
-                dtype=tensor.dtype, count=tensor.numel,
+                dtype=tensor.dtype, count=tensor.numel(),
         ).view(tensor.shape)[:] = tensor
         end = self.size + length
-        self.header[descr] = {
+        self.header[name] = {
             'dtype': 
                 str(tensor.dtype).rsplit('.',1)[-1]
                     .replace('float','F')
@@ -42,14 +53,13 @@ class WritingSafeTensors:
         self.size = end
     def finalize(self):
         print(self.filename, '...')
-        import pdb; pdb.set_trace()
-        header = json.dumps(self.header).encode()
+        header = json.dumps(self.header, separators=[',',':']).encode()
         insert = len(header) + 8
         self._reserve(insert)
         self.mmapview[insert:insert+self.size] = self.mmapview[:self.size]
         self.size += insert
-        self[:8] = len(header).to_bytes(8, 'little')
-        self[8:insert] = header
+        self.mmapview[:8] = len(header).to_bytes(8, 'little')
+        self.mmapview[8:insert] = header
         del self.header
         del self.mmapview
         os.close(self.fd)

run_test.py

@@ -1,5 +1,5 @@
 #!/usr/bin/env python3
-import mmap, os, sys
+import os, sys
 
 STORE_WEIGHTS = False
 FAKE_H100 = False
@@ -75,18 +75,18 @@ def hook(module, inputs, outputs):
         inputs = module._hf_hook.pre_forward(module, *inputs)
     for idx, input in enumerate(inputs):
         if isinstance(input, torch.Tensor):
-            store_tensor(f'{name}.input.{idx}', input);
+            SafeTensors.add(f'{name}.input.{idx}', input);
     if STORE_WEIGHTS and not list(module.children()):
         for wtname, wt in list(module.named_parameters()) + list(module.named_buffers()):
-            store_tensor(f'{name}.{wtname}', wt)
+            SafeTensors.add(f'{name}.{wtname}', wt)
     if HAS_HF_HOOK:
         outputs = module._hf_hook.post_forward(module, outputs)
     if isinstance(outputs, torch.Tensor):
-        store_tensor(f'{name}.output', outputs);
+        SafeTensors.add(f'{name}.output', outputs);
     else:
         for idx, output in enumerate(outputs):
             if isinstance(output, torch.Tensor):
-                store_tensor(f'{name}.output.{idx}', output);
+                SafeTensors.add(f'{name}.output.{idx}', output);
     IDX += 1
 
 for module in pipe.model.modules():