convert : support non-mxfp4 HF model (#15153)

* convert : support non-mxfp4 HF model

* rm redundant check

* disable debug check

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 8e8f09642c4bc2788fe24ebfa732b9fec9aa48a2..b8c7d97a786c759d23b1464f57d9528f1a18f1b3 100755 (executable)
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -8102,7 +8102,6 @@ class GptOssModel(TextModel):
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         blocks0: Tensor = torch.zeros(1)
         blocks1: Tensor = torch.zeros(1)
-        found_mxfp4_tensors = False
         # we assume that tensors are loaded in the correct order
         for name, data_torch in self.get_tensors():
             if "mlp.experts.down_proj_blocks" in name:
@@ -8110,7 +8109,6 @@ class GptOssModel(TextModel):
             elif "mlp.experts.down_proj_scales" in name:
                 new_name = self.map_tensor_name(name.replace("_scales", ".weight"))
                 self.repack_mxfp4(new_name, blocks0, data_torch)
-                found_mxfp4_tensors = True
             elif "mlp.experts.gate_up_proj_blocks" in name:
                 blocks0, blocks1 = data_torch[:, ::2, :, :], data_torch[:, 1::2, :, :]
             elif "mlp.experts.gate_up_proj_scales" in name:
@@ -8119,9 +8117,6 @@ class GptOssModel(TextModel):
                 new_name_up = self.map_tensor_name(name.replace("gate_up_proj_scales", "up_proj.weight"))
                 self.repack_mxfp4(new_name_gate, blocks0, scales0)
                 self.repack_mxfp4(new_name_up, blocks1, scales1)
-                found_mxfp4_tensors = True
-        if not found_mxfp4_tensors:
-            raise ValueError("No MXFP4 tensors found in the model. Please make sure you are using MXFP4 model.")
         return []
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
@@ -8134,7 +8129,12 @@ class GptOssModel(TextModel):
         if "down_proj" in name:
             if name.endswith("_bias"):
                 name = name.replace("down_proj_bias", "down_proj.bias")
+            elif "_blocks" not in name and "_scales" not in name:
+                logger.warning(f"{name} is not in MXFP4, performance may be degraded")
+                name = name.replace("down_proj", "down_proj.weight")
+                data_torch = data_torch.transpose(-1, -2)
             else:
+                # otherwise, it should already be repacked to ggml MXFP4 format
                 return []
 
         # split the gate_up into gate and up
@@ -8147,7 +8147,18 @@ class GptOssModel(TextModel):
                     (self.map_tensor_name(name_gate), gate_proj_bias),
                     (self.map_tensor_name(name_up), up_proj_bias)
                 ]
+            elif "_blocks" not in name and "_scales" not in name:
+                logger.warning(f"{name} is not in MXFP4, performance may be degraded")
+                name_up = name.replace("gate_up_proj", "up_proj.weight")
+                name_gate = name.replace("gate_up_proj", "gate_proj.weight")
+                data_torch = data_torch.transpose(-1, -2)
+                gate_proj_weight, up_proj_weight = data_torch[:, ::2, :], data_torch[:, 1::2, :]
+                return [
+                    (self.map_tensor_name(name_gate), gate_proj_weight),
+                    (self.map_tensor_name(name_up), up_proj_weight)
+                ]
             else:
+                # otherwise, it should already be repacked to ggml MXFP4 format
                 return []
 
         return [(self.map_tensor_name(name), data_torch)]

diff --git a/src/llama-quant.cpp b/src/llama-quant.cpp
index 9c0ffde27a1e3956f2c8db50c78872190050dd75..1d0361cc16659d5d93a5b42c49bb211a52ee4f3d 100644 (file)
--- a/src/llama-quant.cpp
+++ b/src/llama-quant.cpp
@@ -999,7 +999,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
                 new_size += llama_tensor_quantize_impl(new_type, f32_data_03, new_data_03, chunk_size, nrows, n_per_row, imatrix_03, workers, nthread_use);
 
                 // TODO: temporary sanity check that the F16 -> MXFP4 is lossless
-#if 1
+#if 0
                 if (new_type == GGML_TYPE_MXFP4) {
                     auto * x = f32_data_03;