Vulkan IQ4_NL Support (#8613)

* Fix Vulkan matmul tests compile errors

* Add Vulkan IQ4_NL support

* Fix Vulkan DeepSeek-Coder-V2-Lite MoE support

Author: 0cc4m

ggml/src/ggml-vulkan.cpp

@@ -58,3 +58,11 @@ vec2 dequantize(uint ib, uint iqs, uint a_offset) {
     return vec2(int(data_a[a_offset + ib].qs[iqs]), int(data_a[a_offset + ib].qs[iqs + 1])) * d;
 }
 #endif
+
+#if defined(DATA_A_IQ4_NL)
+vec2 dequantize(uint ib, uint iqs, uint a_offset) {
+    const float d = float(data_a[a_offset + ib].d);
+    const uint vui = uint(data_a[a_offset + ib].qs[iqs]);
+    return vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
+}
+#endif

ggml/src/vulkan-shaders/dequant_funcs.comp

@@ -0,0 +1,30 @@
+#version 450
+
+#include "dequant_head.comp"
+
+layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer A {block_iq4_nl data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
+
+void main() {
+    const uint i = gl_WorkGroupID.x * 4 + gl_LocalInvocationID.x / 64;
+
+    const uint tid = gl_LocalInvocationID.x % 64;
+    const uint il  = tid/32;
+    const uint ir  = tid%32;
+    const uint ib = 32*i + ir;
+    if (ib >= p.nel / 32) {
+        return;
+    }
+
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
+
+    const float d = float(data_a[ib].d);
+
+    [[unroll]] for (uint l = 0; l < 8; ++l) {
+        data_b[b_idx + l +  0] = D_TYPE(d * kvalues_iq4nl[data_a[ib].qs[q_idx + l] & 0xF]);
+        data_b[b_idx + l + 16] = D_TYPE(d * kvalues_iq4nl[data_a[ib].qs[q_idx + l] >>  4]);
+    }
+}

ggml/src/vulkan-shaders/dequant_iq4_nl.comp

@@ -18,15 +18,13 @@ void main() {
         return;
     }
 
-    const uint b_idx = 1024*i + 32*ir + 8*il;
+    const uint q_idx = 8*il;
+    const uint b_idx = 1024*i + 32*ir + q_idx;
 
     const float d = float(data_a[ib].d);
-    const float dm = -8.0f * d;
-
-    const uint q_idx = 8*il;
 
     [[unroll]] for (uint l = 0; l < 8; ++l) {
-        data_b[b_idx + l +  0] = D_TYPE(d * (data_a[ib].qs[q_idx + l] & 0xF) + dm);
-        data_b[b_idx + l + 16] = D_TYPE(d * (data_a[ib].qs[q_idx + l] >>  4) + dm);
+        data_b[b_idx + l +  0] = D_TYPE(d * ((data_a[ib].qs[q_idx + l] & 0xF) - 8.0f));
+        data_b[b_idx + l + 16] = D_TYPE(d * ((data_a[ib].qs[q_idx + l] >>  4) - 8.0f));
     }
 }

ggml/src/vulkan-shaders/dequant_q4_0.comp

@@ -71,7 +71,7 @@ shared FLOAT_TYPE buf_a[BM * (BK+1)];
 shared FLOAT_TYPE buf_b[BN * (BK+1)];
 
 #ifdef MUL_MAT_ID
-shared u16vec2 row_ids[2048];
+shared u16vec2 row_ids[3072];
 #endif
 
 void main() {
@@ -380,6 +380,19 @@ void main() {
 
             buf_a[buf_idx    ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi    ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi    ] >> qhshift) & 3) << 4)) - 32));
             buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
+#elif defined(DATA_A_IQ4_NL)
+            const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
+            const uint buf_idx = (loadc_a + l) * (BK+1) + loadr_a;
+
+            const uint ib = idx / 16;
+            const uint iqs = idx & 0xF;
+
+            const float d = float(data_a[ib].d);
+            const uint vui = uint(data_a[ib].qs[iqs]);
+            const vec2 v = vec2(kvalues_iq4nl[vui & 0xF], kvalues_iq4nl[vui >> 4]) * d;
+
+            buf_a[buf_idx     ] = FLOAT_TYPE(v.x);
+            buf_a[buf_idx + 16] = FLOAT_TYPE(v.y);
 #endif
         }
         [[unroll]] for (uint l = 0; l < BN; l += loadstride_b) {

ggml/src/vulkan-shaders/mul_mm.comp

@@ -177,3 +177,24 @@ struct block_q6_K
 
 #define A_TYPE block_q6_K
 #endif
+
+// IQuants
+
+#if defined(DATA_A_IQ4_NL)
+#extension GL_EXT_shader_16bit_storage : require
+#define QUANT_K 32
+#define QUANT_R 2
+
+struct block_iq4_nl
+{
+    float16_t d;
+    uint8_t qs[QUANT_K/2];
+};
+
+#define A_TYPE block_iq4_nl
+
+const int8_t kvalues_iq4nl[16] = {
+    int8_t(-127), int8_t(-104), int8_t(-83), int8_t(-65), int8_t(-49), int8_t(-35), int8_t(-22), int8_t(-10),
+    int8_t(1), int8_t(13), int8_t(25), int8_t(38), int8_t(53), int8_t(69), int8_t(89), int8_t(113)
+};
+#endif

ggml/src/vulkan-shaders/types.comp

@@ -52,7 +52,8 @@ const std::vector<std::string> type_names = {
     "q3_k",
     "q4_k",
     "q5_k",
-    "q6_k"
+    "q6_k",
+    "iq4_nl"
 };
 
 void execute_command(const std::string& command, std::string& stdout_str, std::string& stderr_str) {