llama : add support for BitnetForCausalLM (#7931)

* hf bitnet v1

* hf bitnet e2e v2

* finish bitnet e2e

* finish f16 hf bitnet e2e

* remove unsed

* finish bitnet i2 e2e

* move i2s to quantize v1

* move i2 to quantize

* clean code

* clean code 2

* fix codestyle

* fix code

* fix

* fix code

* fix merge

* remove unused

* change table name

* fix whitespace

* delete redundant

* i2_s to absmax

* finish i2_s/i8_s vec_dot x86 simd

* i2s->q22

* fix code

* remove block scale

* add dequantize

* fix seq

* update avx2

* remove q2_2

* remove q22_grid

* fix whitespace

* reuse llm_build_kv

* fix bo

---------

Co-authored-by: root <root@wangjinheng>

Author: Eddie-Wang1120

convert-hf-to-gguf.py

@@ -1404,6 +1404,48 @@ def write_tensors(self):
                 raise ValueError(f"Unprocessed experts: {experts}")
 
 
+@Model.register("BitnetForCausalLM")
+class BitnetModel(Model):
+    model_arch = gguf.MODEL_ARCH.BITNET
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+        self.gguf_writer.add_rope_scaling_factor(1.0)
+
+    def weight_quant(self, weight):
+        dtype = weight.dtype
+        weight = weight.float()
+        s = 1 / weight.abs().mean().clamp(min=1e-5)
+        weight = (weight * s).round().clamp(-1, 1) / s
+        scale = weight.abs().max().unsqueeze(0)
+        weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
+        weight = torch.sign(weight).type(dtype)
+        return weight.type(dtype), scale.type(torch.float32)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        new_name = self.map_tensor_name(name)
+
+        if any(self.match_model_tensor_name(new_name, key, bid) for key in [
+            gguf.MODEL_TENSOR.ATTN_Q,
+            gguf.MODEL_TENSOR.ATTN_K,
+            gguf.MODEL_TENSOR.ATTN_V,
+            gguf.MODEL_TENSOR.ATTN_OUT,
+            gguf.MODEL_TENSOR.FFN_UP,
+            gguf.MODEL_TENSOR.FFN_DOWN,
+            gguf.MODEL_TENSOR.FFN_GATE,
+        ]):
+            # transform weight into 1/0/-1 (in fp32)
+            weight_torch, scale_torch = self.weight_quant(data_torch)
+            yield (new_name, weight_torch)
+            yield (new_name.removesuffix(".weight") + ".scale", scale_torch)
+        else:
+            yield (new_name, data_torch)
+
+
 @Model.register("GrokForCausalLM")
 class GrokModel(Model):
     model_arch = gguf.MODEL_ARCH.GROK

gguf-py/gguf/constants.py

@@ -149,6 +149,7 @@ class MODEL_ARCH(IntEnum):
     OLMO       = auto()
     ARCTIC     = auto()
     DEEPSEEK2  = auto()
+    BITNET     = auto()
 
 
 class MODEL_TENSOR(IntEnum):
@@ -200,6 +201,8 @@ class MODEL_TENSOR(IntEnum):
     ATTN_KV_B          = auto()
     ATTN_Q_A_NORM      = auto()
     ATTN_KV_A_NORM     = auto()
+    FFN_SUB_NORM       = auto()
+    ATTN_SUB_NORM      = auto()
 
 
 MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
@@ -237,6 +240,7 @@ class MODEL_TENSOR(IntEnum):
     MODEL_ARCH.OLMO:           "olmo",
     MODEL_ARCH.ARCTIC:         "arctic",
     MODEL_ARCH.DEEPSEEK2:      "deepseek2",
+    MODEL_ARCH.BITNET:         "bitnet",
 }
 
 TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
@@ -288,6 +292,8 @@ class MODEL_TENSOR(IntEnum):
     MODEL_TENSOR.ATTN_KV_B:          "blk.{bid}.attn_kv_b",
     MODEL_TENSOR.ATTN_Q_A_NORM:      "blk.{bid}.attn_q_a_norm",
     MODEL_TENSOR.ATTN_KV_A_NORM:     "blk.{bid}.attn_kv_a_norm",
+    MODEL_TENSOR.ATTN_SUB_NORM:      "blk.{bid}.attn_sub_norm",
+    MODEL_TENSOR.FFN_SUB_NORM:       "blk.{bid}.ffn_sub_norm",
 }
 
 MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
@@ -808,6 +814,21 @@ class MODEL_TENSOR(IntEnum):
         MODEL_TENSOR.FFN_DOWN_SHEXP,
         MODEL_TENSOR.FFN_UP_SHEXP,
     ],
+    MODEL_ARCH.BITNET: [
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_SUB_NORM,
+        MODEL_TENSOR.FFN_SUB_NORM,
+    ],
     # TODO
 }
 

gguf-py/gguf/tensor_mapping.py

@@ -413,6 +413,14 @@ class TensorNameMap:
         MODEL_TENSOR.ATTN_KV_A_NORM: (
             "model.layers.{bid}.self_attn.kv_a_layernorm", # deepseek2
         ),
+
+        MODEL_TENSOR.ATTN_SUB_NORM: (
+            "model.layers.{bid}.self_attn.inner_attn_ln",  # bitnet
+        ),
+
+        MODEL_TENSOR.FFN_SUB_NORM: (
+            "model.layers.{bid}.mlp.ffn_layernorm",  # bitnet
+        ),
     }
 
     # architecture-specific block mappings