Mirror torch210 build into torch211 dir (Q5_0/Q5_1 ops)

.gitattributes

@@ -40,3 +40,4 @@ build/torch29-metal-aarch64-darwin/_gguf_dequant_530e374.abi3.so filter=lfs diff
 build/torch210-metal-aarch64-darwin/_gguf_dequant_b5e6652.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch28-metal-aarch64-darwin/_gguf_dequant_b5e6652.abi3.so filter=lfs diff=lfs merge=lfs -text
 build/torch29-metal-aarch64-darwin/_gguf_dequant_b5e6652.abi3.so filter=lfs diff=lfs merge=lfs -text
+build/torch211-metal-aarch64-darwin/_gguf_dequant_b5e6652.abi3.so filter=lfs diff=lfs merge=lfs -text

build/torch211-metal-aarch64-darwin/__init__.py

@@ -72,6 +72,8 @@ def _pipeline(fn_name: str):
 # (kernel_name, bytes_per_block, elems_per_block, threads_per_block_group)
 _FORMATS = {
     "q4_0":   ("dequantize_q4_0",   18, 32, 16),
+    "q5_0":   ("dequantize_q5_0",   22, 32, 16),
+    "q5_1":   ("dequantize_q5_1",   24, 32, 16),
     "q8_0":   ("dequantize_q8_0",   34, 32, 32),
     "q4_K":   ("dequantize_q4_K",  144, 256, 256),
     "q5_K":   ("dequantize_q5_K",  176, 256, 256),
@@ -148,6 +150,8 @@ def dequantize_bytes(fmt: str, quant_bytes) -> np.ndarray:
 # --- Per-format convenience wrappers ----------------------------------------
 
 def dequantize_q4_0_bytes(qb):   return dequantize_bytes("q4_0",   qb)
+def dequantize_q5_0_bytes(qb):   return dequantize_bytes("q5_0",   qb)
+def dequantize_q5_1_bytes(qb):   return dequantize_bytes("q5_1",   qb)
 def dequantize_q8_0_bytes(qb):   return dequantize_bytes("q8_0",   qb)
 def dequantize_q4_K_bytes(qb):   return dequantize_bytes("q4_K",   qb)
 def dequantize_q5_K_bytes(qb):   return dequantize_bytes("q5_K",   qb)
@@ -156,6 +160,8 @@ def dequantize_iq4_nl_bytes(qb): return dequantize_bytes("iq4_nl", qb)
 def dequantize_iq4_xs_bytes(qb): return dequantize_bytes("iq4_xs", qb)
 
 def allocate_q4_0_plan(nblocks):   return allocate_plan("q4_0",   nblocks)
+def allocate_q5_0_plan(nblocks):   return allocate_plan("q5_0",   nblocks)
+def allocate_q5_1_plan(nblocks):   return allocate_plan("q5_1",   nblocks)
 def allocate_q8_0_plan(nblocks):   return allocate_plan("q8_0",   nblocks)
 def allocate_q4_K_plan(nblocks):   return allocate_plan("q4_K",   nblocks)
 def allocate_q5_K_plan(nblocks):   return allocate_plan("q5_K",   nblocks)
@@ -164,6 +170,8 @@ def allocate_iq4_nl_plan(nblocks): return allocate_plan("iq4_nl", nblocks)
 def allocate_iq4_xs_plan(nblocks): return allocate_plan("iq4_xs", nblocks)
 
 def dispatch_q4_0(plan, **kw):   return dispatch(plan, **kw)
+def dispatch_q5_0(plan, **kw):   return dispatch(plan, **kw)
+def dispatch_q5_1(plan, **kw):   return dispatch(plan, **kw)
 def dispatch_q8_0(plan, **kw):   return dispatch(plan, **kw)
 def dispatch_q4_K(plan, **kw):   return dispatch(plan, **kw)
 def dispatch_q5_K(plan, **kw):   return dispatch(plan, **kw)
@@ -196,6 +204,8 @@ def _require_ops():
 
 
 def torch_dequantize_q4_0(input, output):   _require_ops().dequantize_q4_0(input, output)
+def torch_dequantize_q5_0(input, output):   _require_ops().dequantize_q5_0(input, output)
+def torch_dequantize_q5_1(input, output):   _require_ops().dequantize_q5_1(input, output)
 def torch_dequantize_q8_0(input, output):   _require_ops().dequantize_q8_0(input, output)
 def torch_dequantize_q4_K(input, output):   _require_ops().dequantize_q4_K(input, output)
 def torch_dequantize_q5_K(input, output):   _require_ops().dequantize_q5_K(input, output)
@@ -373,6 +383,8 @@ _N_SIMDGROUP = 2  # Q4_0/Q8_0 matvec uses 2 simdgroups per threadgroup
 #   Q6_K       →  2 simdgroups × 1 row    = 2 rows per threadgroup
 _MAT_FORMATS = {
     "q4_0":   (18,  32,  8),
+    "q5_0":   (22,  32,  8),
+    "q5_1":   (24,  32,  8),
     "q8_0":   (34,  32,  8),
     "q4_K":   (144, 256, 4),
     "q5_K":   (176, 256, 4),
@@ -525,7 +537,7 @@ def encode_matvec(plan: MatVecPlan, enc):
     if plan.fmt == "q4_K":
         rows_per_tg = _N_DST
         tg_threads = _SIMD_WIDTH
-    elif plan.fmt in ("q4_0", "q8_0"):
+    elif plan.fmt in ("q4_0", "q5_0", "q5_1", "q8_0"):
         rows_per_tg = _N_SIMDGROUP * _N_DST
         tg_threads = _N_SIMDGROUP * _SIMD_WIDTH
     elif plan.fmt == "q5_K":
@@ -572,7 +584,7 @@ def dispatch_matvec(plan: MatVecPlan, sync: bool = True):
     if plan.fmt == "q4_K":
         rows_per_tg = _N_DST
         tg_threads = _SIMD_WIDTH
-    elif plan.fmt in ("q4_0", "q8_0"):
+    elif plan.fmt in ("q4_0", "q5_0", "q5_1", "q8_0"):
         rows_per_tg = _N_SIMDGROUP * _N_DST
         tg_threads = _N_SIMDGROUP * _SIMD_WIDTH
     elif plan.fmt == "q5_K":

build/torch211-metal-aarch64-darwin/_gguf_dequant_b5e6652.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b63257ac532059de9a913434d4653f070828f0202d42777510c7868bdf4387a0
+size 287104

build/torch211-metal-aarch64-darwin/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _gguf_dequant_530e374
-ops = torch.ops._gguf_dequant_530e374
+from . import _gguf_dequant_b5e6652
+ops = torch.ops._gguf_dequant_b5e6652
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_gguf_dequant_530e374::{op_name}"
+    return f"_gguf_dequant_b5e6652::{op_name}"