Internal change.

BUILD.bazel

@@ -387,6 +387,7 @@ cc_library(
         "ops/matmul_static_nuq.cc",
         "ops/matmul_static_sfp.cc",
         "ops/matmul_static_i8.cc",
+        "ops/matmul_static_q4_0.cc",
     ],
     hdrs = [
         "ops/matmul_static.h",

CMakeLists.txt

@@ -112,6 +112,7 @@ set(SOURCES
   compression/compress.cc
   compression/compress.h
   compression/nuq-inl.h
+  compression/q4_0-inl.h
   compression/sfp-inl.h
   compression/int-inl.h
   compression/types.h
@@ -165,6 +166,7 @@ set(SOURCES
   ops/matmul_static_nuq.cc
   ops/matmul_static_sfp.cc
   ops/matmul_static_i8.cc
+  ops/matmul_static_q4_0.cc
   ops/matmul-inl.h
   ops/matmul.cc
   ops/matmul.h
@@ -295,6 +297,7 @@ set(GEMMA_TEST_FILES
   compression/compress_test.cc
   compression/distortion_test.cc
   compression/nuq_test.cc
+  compression/q4_0_test.cc
   compression/sfp_test.cc
   gemma/gemma_args_test.cc
   gemma/tensor_info_test.cc

compression/BUILD.bazel

@@ -91,6 +91,15 @@ cc_library(
     ],
 )
 
+cc_library(
+    name = "q4_0",
+    textual_hdrs = ["q4_0-inl.h"],
+    deps = [
+        ":types",
+        "@highway//:hwy",
+    ],
+)
+
 cc_test(
     name = "int_test",
     size = "small",
@@ -110,6 +119,26 @@ cc_test(
     ],
 )
 
+cc_test(
+    name = "q4_0_test",
+    size = "small",
+    timeout = "long",
+    srcs = ["q4_0_test.cc"],
+    features = ["fully_static_link"],
+    linkstatic = True,
+    local_defines = ["HWY_IS_TEST"],
+    # for test_suite.
+    tags = ["hwy_ops_test"],
+    deps = [
+        ":compress",
+        ":q4_0",
+        "@googletest//:gtest_main",  # buildcleaner: keep
+        "//:test_util",
+        "@highway//:hwy",
+        "@highway//:hwy_test_util",
+    ],
+)
+
 cc_library(
     name = "test_util",
     textual_hdrs = [
@@ -176,6 +205,7 @@ cc_library(
         ":distortion",
         ":int",
         ":nuq",
+        ":q4_0",
         ":sfp",
         "//:basics",
         "//:mat",

compression/compress-inl.h

@@ -50,6 +50,7 @@
 #include "compression/int-inl.h"
 #include "compression/nuq-inl.h"
 #include "compression/sfp-inl.h"
+#include "compression/q4_0-inl.h"
 
 HWY_BEFORE_NAMESPACE();
 namespace gcpp {
@@ -638,6 +639,39 @@ struct CompressTraits<I8Stream> {
   }
 };
 
+// Q4_0 block quantization.
+template <>
+struct CompressTraits<Q4_0Stream> {
+  using Packed = Q4_0Stream;
+
+  template <class DF, HWY_IF_F32_D(DF)>
+  static HWY_INLINE void Compress(DF df, const float* HWY_RESTRICT raw,
+                                  size_t num, CompressPerThread& tls,
+                                  const PackedSpan<Packed>& packed,
+                                  const size_t packed_ofs) {
+    Q4_0Codec::Enc(df, raw, num, packed, packed_ofs);
+  }
+
+  template <class D>
+  static HWY_INLINE void Load2(D d, const PackedSpan<const Packed>& packed,
+                               const size_t packed_ofs, hn::Vec<D>& raw0,
+                               hn::Vec<D>& raw1) {
+    Q4_0Codec::Dec2(d, packed, packed_ofs, raw0, raw1);
+  }
+
+  static float ToFloatSlow(const Packed x) {
+    HWY_DASSERT(!"Not supported");
+    return 0.0f;
+  }
+
+  template <class D, typename Raw>
+  static HWY_INLINE void DecompressAndZeroPad(
+      D d, const PackedSpan<const Packed>& packed, const size_t packed_ofs,
+      Raw* raw, const size_t num) {
+    Q4_0Codec::DecompressAndZeroPad(d, packed, packed_ofs, raw, num);
+  }
+};
+
 // Nonuniform quantization, 4.5 bits per element, two separate streams.
 template <>
 struct CompressTraits<NuqStream> {

compression/python/compression_clif_aux.cc

@@ -111,6 +111,9 @@ class SbsWriterImpl : public ISbsWriter {
       case Type::kI8:
         InsertT<I8Stream>(name, weights, tensor_info);
         break;
+      case Type::kQ4_0:
+        InsertT<Q4_0Stream>(name, weights, tensor_info);
+        break;
       default:
         HWY_ABORT("Unsupported destination (compressed) type %s",
                   TypeName(type));