LLM inference in C/C++
C++ ggml
convert: prevent ndarray conversion in LazyChunkedTensor (#27869)
Xuan-Son Nguyenb19cbe925be3
diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.pyindex 1f309ad2ea..d95fe9b1ac 100644--- a/gguf-py/gguf/gguf_writer.py+++ b/gguf-py/gguf/gguf_writer.py@@ -467,10 +467,15 @@ class GGUFWriter:shard_bar.reset(total=(total if total > 0 else None))# relying on the fact that Python dicts preserve insertion order (since 3.7)- for ti in tensors.values():+ for name, ti in tensors.items():assert ti.tensor is not None # can only iterate once over the tensorsassert ti.tensor.nbytes == ti.nbytes+ start = fout.tell()ti.tensor.tofile(fout)+ # a short write here would only surface as a corrupt file at load time+ if fout.tell() - start != ti.nbytes:+ raise ValueError(+ f"tensor {name!r} wrote {fout.tell() - start} bytes, expected {ti.nbytes}")if shard_bar is not None:shard_bar.update(ti.nbytes)if bar is not None:diff --git a/gguf-py/gguf/lazy.py b/gguf-py/gguf/lazy.pyindex 6a0aee8811..a39f223215 100644--- a/gguf-py/gguf/lazy.py+++ b/gguf-py/gguf/lazy.py@@ -251,6 +251,10 @@ class LazyChunkedTensor:def numpy(self) -> LazyChunkedTensor:return self+ def __array__(self, *args, **kwargs):+ # numpy would otherwise make a 1-element object array of self, and write 8 bytes+ raise TypeError("LazyChunkedTensor cannot become an ndarray, it is written in chunks")+def quantize(self, qtype: Any) -> LazyChunkedTensor:from .constants import GGMLQuantizationTypefrom .quants import QuantError, quant_shape_to_byte_shape