Merge branch 'master' into compilade/refactor-kv-cache

gguf-py/gguf/constants.py

@@ -1327,6 +1327,8 @@ class GGMLQuantizationType(IntEnum):
     Q4_0_4_4 = 31
     Q4_0_4_8 = 32
     Q4_0_8_8 = 33
+    TQ1_0   = 34
+    TQ2_0   = 35
 
 
 # TODO: add GGMLFileType from ggml_ftype in ggml.h
@@ -1371,6 +1373,8 @@ class LlamaFileType(IntEnum):
     MOSTLY_Q4_0_4_4      = 33  # except 1d tensors
     MOSTLY_Q4_0_4_8      = 34  # except 1d tensors
     MOSTLY_Q4_0_8_8      = 35  # except 1d tensors
+    MOSTLY_TQ1_0         = 36  # except 1d tensors
+    MOSTLY_TQ2_0         = 37  # except 1d tensors
 
     GUESSED              = 1024  # not specified in the model file
 
@@ -1447,6 +1451,8 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
     GGMLQuantizationType.Q4_0_4_4:(32, 2 + 16),
     GGMLQuantizationType.Q4_0_4_8:(32, 2 + 16),
     GGMLQuantizationType.Q4_0_8_8:(32, 2 + 16),
+    GGMLQuantizationType.TQ1_0:   (256, 2 + 4 * 13),
+    GGMLQuantizationType.TQ2_0:   (256, 2 + 64),
 }
 
 

gguf-py/gguf/quants.py

@@ -574,6 +574,87 @@ class Q6_K(__Quant, qtype=GGMLQuantizationType.Q6_K):
         return (d * q).reshape((n_blocks, QK_K))
 
 
+class TQ1_0(__Quant, qtype=GGMLQuantizationType.TQ1_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d = abs(blocks).max(axis=-1, keepdims=True)
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np_roundf(blocks * id)
+        qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8)
+
+        qs0, qs1, qh = qs[..., :(32 * 5)], qs[..., (32 * 5):(48 * 5)], qs[..., (48 * 5):]
+        qs0 = qs0.reshape((n_blocks, -1, 5, 32)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs0 = np.sum(qs0, axis=-2).reshape((n_blocks, -1))
+        qs1 = qs1.reshape((n_blocks, -1, 5, 16)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs1 = np.sum(qs1, axis=-2).reshape((n_blocks, -1))
+        qh = qh.reshape((n_blocks, -1, 4, 4)) * np.array([81, 27, 9, 3], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = np.sum(qh, axis=-2).reshape((n_blocks, -1))
+        qs = np.concatenate([qs0, qs1, qh], axis=-1)
+        qs = (qs.astype(np.uint16) * 256 + (243 - 1)) // 243
+
+        qs = qs.astype(np.uint8)
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([qs, d], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        qs, rest = np.hsplit(blocks, [(QK_K - 4 * QK_K // 64) // 5])
+        qh, d = np.hsplit(rest, [QK_K // 64])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs0, qs1 = qs[..., :32], qs[..., 32:]
+        qs0 = qs0.reshape((n_blocks, -1, 1, 32)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs0 = qs0.reshape((n_blocks, -1))
+        qs1 = qs1.reshape((n_blocks, -1, 1, 16)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1))
+        qs1 = qs1.reshape((n_blocks, -1))
+        qh = qh.reshape((n_blocks, -1, 1, 4)) * np.array([1, 3, 9, 27], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qh = qh.reshape((n_blocks, -1))
+        qs = np.concatenate([qs0, qs1, qh], axis=-1)
+        qs = ((qs.astype(np.uint16) * 3) >> 8).astype(np.int8) - np.int8(1)
+
+        return (d * qs.astype(np.float32))
+
+
+class TQ2_0(__Quant, qtype=GGMLQuantizationType.TQ2_0):
+    @classmethod
+    def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        d = abs(blocks).max(axis=-1, keepdims=True)
+        with np.errstate(divide="ignore"):
+            id = np.where(d == 0, 0, 1 / d)
+        qs = np_roundf(blocks * id)
+        qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8)
+
+        qs = qs.reshape((n_blocks, -1, 4, 32)) << np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qs = qs[..., 0, :] | qs[..., 1, :] | qs[..., 2, :] | qs[..., 3, :]
+        qs = qs.reshape((n_blocks, -1))
+
+        d = d.astype(np.float16).view(np.uint8)
+
+        return np.concatenate([qs, d], axis=-1)
+
+    @classmethod
+    def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
+        n_blocks = blocks.shape[0]
+
+        qs, d = np.hsplit(blocks, [QK_K // 4])
+
+        d = d.view(np.float16).astype(np.float32)
+
+        qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
+        qs = (qs & 0x03).reshape((n_blocks, -1)).astype(np.int8) - np.int8(1)
+
+        return (d * qs.astype(np.float32))
+
+
 class IQ2_XXS(__Quant, qtype=GGMLQuantizationType.IQ2_XXS):
     ksigns: bytes = (
         b"\x00\x81\x82\x03\x84\x05\x06\x87\x88\x09\x0a\x8b\x0c\x8d\x8e\x0f"

gguf-py/pyproject.toml

@@ -23,6 +23,7 @@ python = ">=3.8"
 numpy = ">=1.17"
 tqdm = ">=4.27"
 pyyaml = ">=5.1"
+sentencepiece = ">=0.1.98,<=0.2.0"
 
 [tool.poetry.dev-dependencies]
 pytest = "^5.2"

gguf-py/tests/test_quants.py

@@ -66,6 +66,7 @@ class GGMLQuants:
         for t in (
             "q4_0", "q4_1", "q5_0", "q5_1", "q8_0",
             "q2_K", "q3_K", "q4_K", "q5_K", "q6_K",
+            "tq1_0", "tq2_0",
             "iq2_xxs", "iq2_xs", "iq2_s", "iq3_xxs", "iq3_s", "iq1_s", "iq1_m",
             "iq4_nl", "iq4_xs",
         ):