[dflash] Add BF16 support to GGUF loader (#20705)

metascroy · web-flow · commit 572f102bc7d0 · 2026-07-02T21:12:55.000-07:00
diff --git a/extension/llm/export/gguf.py b/extension/llm/export/gguf.py
@@ -380,6 +380,8 @@ def iter_gguf(
             yield tensor.name, flat.view(torch.float16).reshape(shape).to(
                 torch.bfloat16
             )
+        elif tensor.tensor_type == GGMLQuantizationType.BF16:
+            yield tensor.name, flat.view(torch.bfloat16).reshape(shape).clone()
         else:
             raise ValueError(f"Unsupported GGUF quant type: {tensor.tensor_type}")
 

Original file line number	Diff line number	Diff line change
`@@ -380,6 +380,8 @@ def iter_gguf(`
`380`	`380`	`yield tensor.name, flat.view(torch.float16).reshape(shape).to(`
`381`	`381`	`torch.bfloat16`
`382`	`382`	`)`
	`383`	`+ elif tensor.tensor_type == GGMLQuantizationType.BF16:`
	`384`	`+ yield tensor.name, flat.view(torch.bfloat16).reshape(shape).clone()`
`383`	`385`	`else:`
`384`	`386`	`raise ValueError(f"Unsupported GGUF quant type: {tensor.tensor_type}")`
`385`	`387`