whispercpp 1.3.5 → 1.3.7

data/ext/sources/ggml/src/ggml-cuda/mmvq.cu

@@ -9,12 +9,14 @@ typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_
 
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return vec_dot_q1_0_q8_1;
         case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
         case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
         case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
         case GGML_TYPE_Q5_1:    return vec_dot_q5_1_q8_1;
         case GGML_TYPE_Q8_0:    return vec_dot_q8_0_q8_1;
         case GGML_TYPE_MXFP4:   return vec_dot_mxfp4_q8_1;
+        case GGML_TYPE_NVFP4:   return vec_dot_nvfp4_q8_1;
         case GGML_TYPE_Q2_K:    return vec_dot_q2_K_q8_1;
         case GGML_TYPE_Q3_K:    return vec_dot_q3_K_q8_1;
         case GGML_TYPE_Q4_K:    return vec_dot_q4_K_q8_1;
@@ -33,14 +35,16 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
     }
 }
 
-static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
+static constexpr __host__ __device__ int get_vdr_mmvq(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return VDR_Q1_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
         case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
         case GGML_TYPE_Q5_1:    return VDR_Q5_1_Q8_1_MMVQ;
         case GGML_TYPE_Q8_0:    return VDR_Q8_0_Q8_1_MMVQ;
         case GGML_TYPE_MXFP4:   return VDR_MXFP4_Q8_1_MMVQ;
+        case GGML_TYPE_NVFP4:   return VDR_NVFP4_Q8_1_MMVQ;
         case GGML_TYPE_Q2_K:    return VDR_Q2_K_Q8_1_MMVQ;
         case GGML_TYPE_Q3_K:    return VDR_Q3_K_Q8_1_MMVQ;
         case GGML_TYPE_Q4_K:    return VDR_Q4_K_Q8_1_MMVQ;
@@ -59,31 +63,290 @@ static constexpr __device__ int get_vdr_mmvq(ggml_type type) {
 
 enum mmvq_parameter_table_id {
     MMVQ_PARAMETERS_GENERIC = 0,
+    MMVQ_PARAMETERS_TURING,
     MMVQ_PARAMETERS_GCN,
-    MMVQ_PARAMETERS_RDNA2
+    MMVQ_PARAMETERS_RDNA2,
+    MMVQ_PARAMETERS_RDNA3_0,
+    MMVQ_PARAMETERS_RDNA4
 };
 
 static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
-#if defined(RDNA2) || defined(RDNA3) || defined(RDNA4)
+#if defined(RDNA4)
+    return MMVQ_PARAMETERS_RDNA4;
+#elif defined(RDNA3_0)
+    return MMVQ_PARAMETERS_RDNA3_0;
+#elif defined(RDNA2) || defined(RDNA3_5)
     return MMVQ_PARAMETERS_RDNA2;
 #elif defined(GCN) || defined(CDNA)
     return MMVQ_PARAMETERS_GCN;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING && __CUDA_ARCH__ < GGML_CUDA_CC_AMPERE
+    return MMVQ_PARAMETERS_TURING;
 #else
     return MMVQ_PARAMETERS_GENERIC;
 #endif
 }
 
 static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
-    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+    if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+        return MMVQ_PARAMETERS_RDNA4;
+    }
+    if (GGML_CUDA_CC_IS_RDNA3_0(cc)) {
+        return MMVQ_PARAMETERS_RDNA3_0;
+    }
+    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc)) {
         return MMVQ_PARAMETERS_RDNA2;
     }
     if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
         return MMVQ_PARAMETERS_GCN;
     }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_TURING && ggml_cuda_highest_compiled_arch(cc) < GGML_CUDA_CC_AMPERE) {
+        return MMVQ_PARAMETERS_TURING;
+    }
     return MMVQ_PARAMETERS_GENERIC;
 }
 
-static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_parameter_table_id table_id) {
+// Per-architecture maximum batch size for which MMVQ should be used for MUL_MAT_ID.
+// Returns a value <= MMVQ_MAX_BATCH_SIZE. Default is MMVQ_MAX_BATCH_SIZE.
+// Check https://github.com/ggml-org/llama.cpp/pull/20905#issuecomment-4145835627 for details
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_pascal_older(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 4;
+        case GGML_TYPE_NVFP4:   return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 6;
+        case GGML_TYPE_Q4_1:    return 6;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_0:    return 6;
+        case GGML_TYPE_Q5_1:    return 6;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_turing_plus(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 7;
+        case GGML_TYPE_IQ3_S:   return 6;
+        case GGML_TYPE_IQ3_XXS: return 7;
+        case GGML_TYPE_MXFP4:   return 7;
+        case GGML_TYPE_NVFP4:   return 8;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_gcn(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 5;
+        case GGML_TYPE_IQ1_M:   return 5;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 4;
+        case GGML_TYPE_Q2_K:    return 4;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 5;
+        case GGML_TYPE_Q4_1:    return 5;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        case GGML_TYPE_Q8_0:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_cdna(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 5;
+        case GGML_TYPE_IQ2_XS:  return 5;
+        case GGML_TYPE_IQ2_XXS: return 5;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna1_rdna2(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_Q2_K:    return 7;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_K:    return 5;
+        case GGML_TYPE_Q5_K:    return 6;
+        case GGML_TYPE_Q6_K:    return 5;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna3(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 6;
+        case GGML_TYPE_IQ1_M:   return 6;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 6;
+        case GGML_TYPE_IQ4_XS:  return 6;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_K:    return 4;
+        case GGML_TYPE_Q6_K:    return 4;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+static constexpr __host__ __device__ int get_mmvq_mmid_max_batch_rdna4(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_IQ1_S:   return 7;
+        case GGML_TYPE_IQ1_M:   return 7;
+        case GGML_TYPE_IQ2_S:   return 4;
+        case GGML_TYPE_IQ2_XS:  return 4;
+        case GGML_TYPE_IQ2_XXS: return 4;
+        case GGML_TYPE_IQ3_S:   return 4;
+        case GGML_TYPE_IQ3_XXS: return 4;
+        case GGML_TYPE_IQ4_NL:  return 7;
+        case GGML_TYPE_IQ4_XS:  return 5;
+        case GGML_TYPE_MXFP4:   return 5;
+        case GGML_TYPE_NVFP4:   return 5;
+        case GGML_TYPE_Q3_K:    return 4;
+        case GGML_TYPE_Q4_0:    return 7;
+        case GGML_TYPE_Q4_1:    return 7;
+        case GGML_TYPE_Q4_K:    return 4;
+        case GGML_TYPE_Q5_0:    return 7;
+        case GGML_TYPE_Q5_1:    return 7;
+        case GGML_TYPE_Q5_K:    return 5;
+        case GGML_TYPE_Q6_K:    return 5;
+        case GGML_TYPE_Q8_0:    return 7;
+        default:                return MMVQ_MAX_BATCH_SIZE;
+    }
+}
+
+// Host function: returns the max batch size for the current arch+type at runtime.
+int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
+    // NVIDIA: Volta, Ada Lovelace, and Blackwell always use MMVQ for MUL_MAT_ID.
+    if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+        if (cc == GGML_CUDA_CC_VOLTA || cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+            return MMVQ_MAX_BATCH_SIZE;
+        }
+        if (cc >= GGML_CUDA_CC_TURING) {
+            return get_mmvq_mmid_max_batch_turing_plus(type);
+        }
+        return get_mmvq_mmid_max_batch_pascal_older(type);
+    }
+
+    // AMD
+    if (GGML_CUDA_CC_IS_AMD(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+            return get_mmvq_mmid_max_batch_rdna4(type);
+        }
+        if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+            return get_mmvq_mmid_max_batch_rdna3(type);
+        }
+        if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
+            return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+        }
+        if (GGML_CUDA_CC_IS_CDNA(cc)) {
+            return get_mmvq_mmid_max_batch_cdna(type);
+        }
+        if (GGML_CUDA_CC_IS_GCN(cc)) {
+            return get_mmvq_mmid_max_batch_gcn(type);
+        }
+    }
+    return MMVQ_MAX_BATCH_SIZE;
+}
+
+bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
+    if (GGML_CUDA_CC_IS_CDNA(cc)) {
+        if (GGML_CUDA_CC_IS_CDNA1(cc)) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                    return ne11 <= 7;
+                case GGML_TYPE_Q5_1:
+                    return ne11 <= 7;
+                case GGML_TYPE_Q8_0:
+                    return ne11 <= 6;
+                case GGML_TYPE_Q2_K:
+                    return ne11 <= 4;
+                case GGML_TYPE_Q3_K:
+                    return ne11 <= 3;
+                case GGML_TYPE_Q4_K:
+                    return ne11 <= 2;
+                case GGML_TYPE_Q5_K:
+                    return ne11 <= 3;
+                case GGML_TYPE_Q6_K:
+                    return ne11 <= 4;
+                case GGML_TYPE_IQ1_S:
+                    return ne11 <= 5;
+                case GGML_TYPE_IQ2_XXS:
+                case GGML_TYPE_IQ3_S:
+                case GGML_TYPE_IQ4_XS:
+                    return ne11 <= 6;
+                default:
+                    return ne11 <= MMVQ_MAX_BATCH_SIZE;
+            }
+        }
+        switch (type) { // tuned for CDNA2
+            case GGML_TYPE_Q2_K:
+                return ne11 <= 5;
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+                return ne11 <= 3;
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 5;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
+    return ne11 <= MMVQ_MAX_BATCH_SIZE;
+}
+
+// Device constexpr: returns the max batch size for the current arch+type at compile time.
+template <ggml_type type>
+static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
+#if defined(RDNA4)
+    return get_mmvq_mmid_max_batch_rdna4(type);
+#elif defined(RDNA3)
+    return get_mmvq_mmid_max_batch_rdna3(type);
+#elif defined(RDNA2) || defined(RDNA1)
+    return get_mmvq_mmid_max_batch_rdna1_rdna2(type);
+#elif defined(CDNA)
+    return get_mmvq_mmid_max_batch_cdna(type);
+#elif defined(GCN)
+    return get_mmvq_mmid_max_batch_gcn(type);
+#elif defined(__CUDA_ARCH__) && (__CUDA_ARCH__ == GGML_CUDA_CC_VOLTA || __CUDA_ARCH__ >= GGML_CUDA_CC_ADA_LOVELACE)
+    return MMVQ_MAX_BATCH_SIZE;
+#elif defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= GGML_CUDA_CC_TURING
+    return get_mmvq_mmid_max_batch_turing_plus(type);
+#else
+    return get_mmvq_mmid_max_batch_pascal_older(type);
+#endif
+}
+
+static constexpr __host__ __device__ int calc_nwarps(ggml_type type, int ncols_dst, mmvq_parameter_table_id table_id) {
     if (table_id == MMVQ_PARAMETERS_GENERIC) {
         switch (ncols_dst) {
             case 1:
@@ -114,14 +377,86 @@ static constexpr __host__ __device__ int calc_nwarps(int ncols_dst, mmvq_paramet
                 return 1;
         }
     }
+    if (table_id == MMVQ_PARAMETERS_RDNA4) {
+        // nwarps=8 benefits types with simple vec_dot on RDNA4 (ncols_dst=1).
+        // Types with complex vec_dot (Q3_K, IQ2_*, IQ3_*) regress due to register
+        // pressure and lookup table contention at higher thread counts.
+        if (ncols_dst == 1) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                case GGML_TYPE_IQ4_NL:
+                case GGML_TYPE_IQ4_XS:
+                    return 8;
+                default:
+                    return 1;
+            }
+        }
+        return 1;
+    }
+    if (table_id == MMVQ_PARAMETERS_RDNA3_0) {
+        // RDNA3 (W7900): stricter whitelist than RDNA4.
+        // Q2_K / Q5_K / IQ4_XS regress in full quant sweeps.
+        if (ncols_dst == 1) {
+            switch (type) {
+                case GGML_TYPE_Q4_0:
+                case GGML_TYPE_Q4_1:
+                case GGML_TYPE_Q5_0:
+                case GGML_TYPE_Q5_1:
+                case GGML_TYPE_Q8_0:
+                    return 8;
+                case GGML_TYPE_Q6_K:
+                    return 2;
+                case GGML_TYPE_IQ4_NL:
+                    return 8;
+                default:
+                    return 1;
+            }
+        }
+        return 1;
+    }
+    if (table_id == MMVQ_PARAMETERS_TURING) {
+        if (ncols_dst == 1) {
+            switch (type) {
+                case GGML_TYPE_Q2_K:
+                case GGML_TYPE_Q3_K:
+                case GGML_TYPE_Q4_K:
+                case GGML_TYPE_Q5_K:
+                case GGML_TYPE_Q6_K:
+                    return 2;
+                default:
+                    return 4;
+            }
+        }
+        switch (ncols_dst) {
+            case 2:
+            case 3:
+            case 4:
+                return 4;
+            case 5:
+            case 6:
+            case 7:
+            case 8:
+                return 2;
+            default:
+                return 1;
+        }
+    }
     return 1;
 }
 
-static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id) {
-    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN) {
+static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int table_id, bool small_k = false, int nwarps = 1) {
+    if (table_id == MMVQ_PARAMETERS_GENERIC || table_id == MMVQ_PARAMETERS_GCN || table_id == MMVQ_PARAMETERS_TURING) {
         switch (ncols_dst) {
             case 1:
-                return 1;
+                return small_k ? nwarps : 1;
             case 2:
             case 3:
             case 4:
@@ -137,22 +472,26 @@ static constexpr __host__ __device__ int calc_rows_per_block(int ncols_dst, int
     return 1;
 }
 
-// tell the compiler to use as many registers as it wants, see nwarps definition below
-template <ggml_type type, int ncols_dst, bool has_fusion>
-__launch_bounds__(calc_nwarps(ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
+template <ggml_type type, int ncols_dst, bool has_fusion, bool small_k = false>
+__launch_bounds__(calc_nwarps(type, ncols_dst, get_device_table_id())*ggml_cuda_get_physical_warp_size(), 1)
 static __global__ void mul_mat_vec_q(
-        const void * __restrict__ vx, const void * __restrict__ vy, const int32_t * __restrict__ ids, const ggml_cuda_mm_fusion_args_device fusion, float * __restrict__ dst,
+        const void * vx_ptr, const void * vy_ptr, const int32_t * ids_ptr, const ggml_cuda_mm_fusion_args_device fusion, float * dst_ptr,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
-        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst) {
+        const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
+        const uint32_t ids_stride) {
+    const void    * GGML_CUDA_RESTRICT vx  = vx_ptr;
+    const void    * GGML_CUDA_RESTRICT vy  = vy_ptr;
+    const int32_t * GGML_CUDA_RESTRICT ids = ids_ptr;
+    float         * GGML_CUDA_RESTRICT dst = dst_ptr;
 
     constexpr int qk  = ggml_cuda_type_traits<type>::qk;
     constexpr int qi  = ggml_cuda_type_traits<type>::qi;
     constexpr int vdr = get_vdr_mmvq(type);
     constexpr mmvq_parameter_table_id table_id = get_device_table_id();
-    constexpr int nwarps = calc_nwarps(ncols_dst, table_id);
-    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id);
+    constexpr int nwarps = calc_nwarps(type, ncols_dst, table_id);
+    constexpr int rows_per_cuda_block = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
 
     constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
@@ -162,18 +501,24 @@ static __global__ void mul_mat_vec_q(
     const     int blocks_per_row_x = ncols_x / qk;
     constexpr int blocks_per_iter = vdr * nwarps*warp_size / qi;
 
-    // The MUL_MAT_ID code path with ids != nullptr is only implemented for ncols_dst == 1.
     const uint32_t channel_dst = blockIdx.y;
-    const uint32_t channel_x   = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
-    const uint32_t channel_y   = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
-    const uint32_t sample_dst  = blockIdx.z;
+
+    uint32_t channel_x;
+    uint32_t channel_y;
+    uint32_t sample_dst;
+
+    ggml_cuda_pdl_sync();
+    channel_x  = ncols_dst == 1 && ids ? ids[channel_dst]                     : fastdiv(channel_dst, channel_ratio);
+    channel_y  = ncols_dst == 1 && ids ? fastmodulo(channel_dst, nchannels_y) : channel_dst;
+    sample_dst = blockIdx.z;
+
     const uint32_t sample_x    = fastdiv(sample_dst, sample_ratio);
     const uint32_t sample_y    = sample_dst;
 
     bool use_gate = false;
     bool use_bias = false;
     bool use_gate_bias = false;
-    const void * vgate = nullptr;
+    [[maybe_unused]] const void * vgate = nullptr;
     const float * x_bias = nullptr;
     const float * gate_bias = nullptr;
     ggml_glu_op active_glu;
@@ -188,11 +533,11 @@ static __global__ void mul_mat_vec_q(
         active_glu    = fusion.glu_op;
     }
 
-    const uint32_t channel_bias = ids ? channel_x : channel_dst;
 
-    float x_biases[ncols_dst]    = { 0.0f };
-    float gate_biases[ncols_dst] = { 0.0f };
+    [[maybe_unused]] float x_biases[ncols_dst]    = { 0.0f };
+    [[maybe_unused]] float gate_biases[ncols_dst] = { 0.0f };
     if constexpr (has_fusion) {
+        const uint32_t channel_bias = ids ? channel_x : channel_dst;
         if (use_bias) {
             x_bias = x_bias + sample_dst*stride_sample_dst + channel_bias*stride_channel_dst + row0;
             // 1. Hide latency by prefetching bias and gate here
@@ -247,12 +592,7 @@ static __global__ void mul_mat_vec_q(
     }
 
     __shared__ float tmp_shared[nwarps-1 > 0 ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
-    __shared__ float tmp_shared_gate[(has_fusion && (nwarps-1 > 0)) ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
-    if constexpr (!has_fusion) {
-        (void) tmp_shared_gate;
-    } else if (!use_gate) {
-        (void) tmp_shared_gate;
-    }
+    [[maybe_unused]] __shared__ float tmp_shared_gate[(has_fusion && (nwarps-1 > 0)) ? nwarps-1 : 1][ncols_dst][rows_per_cuda_block][warp_size];
 
     if (threadIdx.y > 0) {
 #pragma unroll
@@ -334,41 +674,139 @@ static __global__ void mul_mat_vec_q(
     }
 }
 
+// Dedicated MoE multi-token kernel.
+// Grid: (ceil(nrows_x / c_rows_per_block), nchannels_dst)
+// Block: (warp_size, ncols_dst) - each warp handles one token independently.
+// No shared memory reduction needed since each warp works alone.
+template <ggml_type type, int c_rows_per_block>
+__launch_bounds__(get_mmvq_mmid_max_batch_for_device<type>()*ggml_cuda_get_physical_warp_size(), 1)
+static __global__ void mul_mat_vec_q_moe(
+        const void * vx_ptr, const void * vy_ptr, const int32_t * ids_ptr,
+        float * dst_ptr,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride) {
+    const void    * GGML_CUDA_RESTRICT vx  = vx_ptr;
+    const void    * GGML_CUDA_RESTRICT vy  = vy_ptr;
+    const int32_t * GGML_CUDA_RESTRICT ids = ids_ptr;
+    float         * GGML_CUDA_RESTRICT dst = dst_ptr;
+
+    constexpr int qk  = ggml_cuda_type_traits<type>::qk;
+    constexpr int qi  = ggml_cuda_type_traits<type>::qi;
+    constexpr int vdr = get_vdr_mmvq(type);
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+    constexpr vec_dot_q_cuda_t vec_dot_q_cuda = get_vec_dot_q_cuda(type);
+
+    const uint32_t token_idx   = threadIdx.y;
+    const int      row0        = c_rows_per_block*blockIdx.x;
+    const int      blocks_per_row_x = ncols_x / qk;
+    constexpr int  blocks_per_iter  = vdr * warp_size / qi;
+
+    const uint32_t channel_dst = blockIdx.y;
+
+    if (token_idx >= ncols_dst) {
+        return;
+    }
+
+    ggml_cuda_pdl_sync();
+    const uint32_t channel_x = ids[channel_dst + token_idx * ids_stride];
+    const uint32_t channel_y = fastmodulo(channel_dst, nchannels_y);
+
+    const block_q8_1 * y = ((const block_q8_1 *) vy) + channel_y*stride_channel_y + token_idx*stride_col_y;
+    const int kbx_offset  = channel_x*stride_channel_x + row0*stride_row_x;
+
+    // partial sum for each thread
+    float tmp[c_rows_per_block] = {0.0f};
+
+    for (int kbx = threadIdx.x / (qi/vdr); kbx < blocks_per_row_x; kbx += blocks_per_iter) {
+        const int kby = kbx * (qk/QK8_1);
+        const int kqs = vdr * (threadIdx.x % (qi/vdr));
+
+#pragma unroll
+        for (int i = 0; i < c_rows_per_block; ++i) {
+            tmp[i] += vec_dot_q_cuda(vx, &y[kby], kbx_offset + i*stride_row_x + kbx, kqs);
+        }
+    }
+
+    ggml_cuda_pdl_lc();
+
+    // Warp-level reduction only - no shared memory needed
+#pragma unroll
+    for (int i = 0; i < c_rows_per_block; ++i) {
+        tmp[i] = warp_reduce_sum<warp_size>(tmp[i]);
+    }
+
+    // Write results
+    if (threadIdx.x < c_rows_per_block && (c_rows_per_block == 1 || uint32_t(row0 + threadIdx.x) < nrows_x)) {
+        dst[channel_dst*stride_channel_dst + token_idx*stride_col_dst + row0 + threadIdx.x] = tmp[threadIdx.x];
+    }
+}
+
+template<ggml_type type>
 static std::pair<dim3, dim3> calc_launch_params(
-        const int ncols_dst, const int nrows_x, const int nchannels_y, const int nsamples_y,
-        const int warp_size, const mmvq_parameter_table_id table_id) {
-    const int64_t nblocks = (nrows_x + calc_rows_per_block(ncols_dst, table_id) - 1) / calc_rows_per_block(ncols_dst, table_id);
-    const dim3 block_nums(nblocks, nchannels_y, nsamples_y);
-    const dim3 block_dims(warp_size, calc_nwarps(ncols_dst, table_id), 1);
+        const int ncols_dst, const int nrows_x, const int nchannels_dst, const int nsamples_or_ntokens,
+        const int warp_size, const mmvq_parameter_table_id table_id, const bool small_k = false) {
+    const int nwarps = calc_nwarps(type, ncols_dst, table_id);
+    const int rpb = calc_rows_per_block(ncols_dst, table_id, small_k, nwarps);
+    const int64_t nblocks = (nrows_x + rpb - 1) / rpb;
+    const dim3 block_nums(nblocks, nchannels_dst, nsamples_or_ntokens);
+    const dim3 block_dims(warp_size, nwarps, 1);
     return {block_nums, block_dims};
 }
 
-template<ggml_type type, int c_ncols_dst>
+template<ggml_type type, int c_ncols_dst, bool small_k = false>
 static void mul_mat_vec_q_switch_fusion(
         const void * vx, const void * vy, const int32_t * ids, const ggml_cuda_mm_fusion_args_device fusion, float * dst,
         const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t stride_row_x, const uint32_t stride_col_y,
         const uint32_t stride_col_dst, const uint3 channel_ratio, const uint32_t stride_channel_x,
         const uint32_t stride_channel_y, const uint32_t stride_channel_dst, const uint3 sample_ratio,
         const uint32_t stride_sample_x, const uint32_t stride_sample_y, const uint32_t stride_sample_dst,
-        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared, cudaStream_t stream) {
+        const dim3 & block_nums, const dim3 & block_dims, const int nbytes_shared,
+        const uint32_t ids_stride, cudaStream_t stream) {
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
     if constexpr (c_ncols_dst == 1) {
         if (has_fusion) {
-            mul_mat_vec_q<type, c_ncols_dst, true><<<block_nums, block_dims, nbytes_shared, stream>>>
-                (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+            const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
+            ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, true, small_k>, launch_params,
+                 vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+                 sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
             return;
         }
     }
 
     GGML_ASSERT(!has_fusion && "fusion only supported for ncols_dst=1");
 
-    mul_mat_vec_q<type, c_ncols_dst, false><<<block_nums, block_dims, nbytes_shared, stream>>>
-        (vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, nbytes_shared, stream);
+    ggml_cuda_kernel_launch(mul_mat_vec_q<type, c_ncols_dst, false, small_k>, launch_params,
+        vx, vy, ids, fusion, dst, ncols_x, nchannels_y, stride_row_x, stride_col_y, stride_col_dst,
         channel_ratio, stride_channel_x, stride_channel_y, stride_channel_dst,
-        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst);
+        sample_ratio, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride);
+}
+
+template <ggml_type type>
+static void mul_mat_vec_q_moe_launch(
+        const void * vx, const void * vy, const int32_t * ids, float * dst,
+        const uint32_t ncols_x, const uint3 nchannels_y, const uint32_t nrows_x,
+        const uint32_t stride_row_x, const uint32_t stride_col_y, const uint32_t stride_col_dst,
+        const uint32_t stride_channel_x, const uint32_t stride_channel_y, const uint32_t stride_channel_dst,
+        const uint32_t ncols_dst, const uint32_t ids_stride,
+        const int warp_size, const int nchannels_dst, cudaStream_t stream) {
+
+    constexpr int rows_per_block = 2; // 2 gives best perf based on tuning
+    const int64_t nblocks_rows = (nrows_x + rows_per_block - 1) / rows_per_block;
+    const dim3 block_nums(nblocks_rows, nchannels_dst);
+    const dim3 block_dims(warp_size, ncols_dst);
+    const ggml_cuda_kernel_launch_params launch_params = ggml_cuda_kernel_launch_params(block_nums, block_dims, 0, stream);
+
+    ggml_cuda_kernel_launch(mul_mat_vec_q_moe<type, rows_per_block>, launch_params,
+        vx, vy, ids, dst, ncols_x, nchannels_y, nrows_x,
+        stride_row_x, stride_col_y, stride_col_dst,
+        stride_channel_x, stride_channel_y, stride_channel_dst,
+        ncols_dst, ids_stride);
 }
 
 template <ggml_type type>
@@ -379,7 +817,7 @@ static void mul_mat_vec_q_switch_ncols_dst(
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
         const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
 
     GGML_ASSERT(ncols_x % ggml_blck_size(type) == 0);
     GGML_ASSERT(ncols_dst <= MMVQ_MAX_BATCH_SIZE);
@@ -389,76 +827,144 @@ static void mul_mat_vec_q_switch_ncols_dst(
     const uint3 sample_ratio_fd  = init_fastdiv_values(nsamples_dst  / nsamples_x);
 
     const int device = ggml_cuda_get_device();
+    const int                     cc        = ggml_cuda_info().devices[device].cc;
     const int warp_size = ggml_cuda_info().devices[device].warp_size;
-    const mmvq_parameter_table_id table_id = get_device_table_id(ggml_cuda_info().devices[device].cc);
+    const mmvq_parameter_table_id table_id  = get_device_table_id(cc);
 
     const bool has_fusion = fusion.gate != nullptr || fusion.x_bias != nullptr || fusion.gate_bias != nullptr;
+    const bool has_ids = ids != nullptr;
+
+    const auto should_use_small_k = [&](int c_ncols_dst) {
+        // When K is small, increase rows_per_block to match nwarps so each warp has more work to do
+        // Trigger when the full thread block covers all K blocks in a single loop iteration and few threads remain idle.
+        constexpr int qk                    = ggml_cuda_type_traits<type>::qk;
+        constexpr int qi                    = ggml_cuda_type_traits<type>::qi;
+        constexpr int vdr                   = get_vdr_mmvq(type);
+        const int     blocks_per_row_x      = ncols_x / qk;
+        const int     blocks_per_iter_1warp = vdr * warp_size / qi;
+        const int     nwarps                = calc_nwarps(type, c_ncols_dst, table_id);
+        bool          use                   = nwarps > 1 && blocks_per_row_x < nwarps * blocks_per_iter_1warp;
+
+        constexpr std::array<ggml_type, 2> iq_slow_turing = {
+            GGML_TYPE_IQ3_XXS,
+            GGML_TYPE_IQ3_S,
+        };
+        constexpr std::array<ggml_type, 8> iq_slow_other = {
+            GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M,   GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS,
+            GGML_TYPE_IQ2_S, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S,   GGML_TYPE_IQ4_XS,
+        };
+        constexpr std::array<ggml_type, 3> slow_pascal = {
+            GGML_TYPE_IQ3_S,
+            GGML_TYPE_Q2_K,
+            GGML_TYPE_Q3_K,
+        };
+
+        const bool is_nvidia_turing_plus  = GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_TURING;
+        const bool is_nvidia_pascal_older = GGML_CUDA_CC_IS_NVIDIA(cc) && cc < GGML_CUDA_CC_VOLTA;
+
+        if (is_nvidia_turing_plus) {
+            if (ncols_dst == 1 &&
+                    std::find(iq_slow_turing.begin(), iq_slow_turing.end(), type) != iq_slow_turing.end()) {
+                use = false;
+            }
+        } else if ((ncols_dst == 1 && std::find(iq_slow_other.begin(), iq_slow_other.end(), type) != iq_slow_other.end()) ||
+                (is_nvidia_pascal_older && std::find(slow_pascal.begin(), slow_pascal.end(), type) != slow_pascal.end()) ||
+                GGML_CUDA_CC_IS_RDNA(cc)) {
+            use = false;
+        }
+
+        return use;
+    };
+
+    if (has_ids && ncols_dst > 1) {
+        // Multi-token MUL_MAT_ID path - dedicated MoE kernel
+        mul_mat_vec_q_moe_launch<type>(
+            vx, vy, ids, dst, ncols_x, nchannels_y_fd, nrows_x,
+            stride_row_x, stride_col_y, stride_col_dst,
+            stride_channel_x, stride_channel_y, stride_channel_dst,
+            ncols_dst, ids_stride, warp_size, nchannels_dst, stream);
+        return;
+    }
 
-    GGML_ASSERT(!ids || ncols_dst == 1);
     switch (ncols_dst) {
         case 1: {
             constexpr int c_ncols_dst = 1;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
-            mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
-                 channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+
+            bool use_small_k = should_use_small_k(c_ncols_dst);
+
+            if (use_small_k) {
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id, true);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst, true>(
+                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
+            } else {
+                std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst,
+                                                                        nsamples_dst, warp_size, table_id);
+                mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(
+                    vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
+                    channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst, sample_ratio_fd,
+                    stride_sample_x, stride_sample_y, stride_sample_dst, dims.first, dims.second, 0, ids_stride,
+                    stream);
+            }
         } break;
         case 2: {
             constexpr int c_ncols_dst = 2;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 3: {
             constexpr int c_ncols_dst = 3;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 4: {
             constexpr int c_ncols_dst = 4;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 5: {
             constexpr int c_ncols_dst = 5;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 6: {
             constexpr int c_ncols_dst = 6;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 7: {
             constexpr int c_ncols_dst = 7;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         case 8: {
             constexpr int c_ncols_dst = 8;
-            std::pair<dim3, dim3> dims = calc_launch_params(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
+            std::pair<dim3, dim3> dims = calc_launch_params<type>(c_ncols_dst, nrows_x, nchannels_dst, nsamples_dst, warp_size, table_id);
             mul_mat_vec_q_switch_fusion<type, c_ncols_dst>(vx, vy, ids, fusion, dst, ncols_x, nchannels_y_fd, stride_row_x, stride_col_y, stride_col_dst,
                  channel_ratio_fd, stride_channel_x, stride_channel_y, stride_channel_dst,
                  sample_ratio_fd, stride_sample_x, stride_sample_y, stride_sample_dst,
-                 dims.first, dims.second, 0, stream);
+                 dims.first, dims.second, 0, ids_stride, stream);
         } break;
         default:
             GGML_ABORT("fatal error");
@@ -474,127 +980,139 @@ static void mul_mat_vec_q_switch_type(
         const int nchannels_x, const int nchannels_y, const int nchannels_dst,
         const int stride_channel_x, const int stride_channel_y, const int stride_channel_dst,
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
-        cudaStream_t stream) {
+        const int ids_stride, cudaStream_t stream) {
     switch (type_x) {
+        case GGML_TYPE_Q1_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q1_0>
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
+            break;
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q4_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_1>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_1:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_1>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q8_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q8_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_MXFP4:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_MXFP4>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
+            break;
+        case GGML_TYPE_NVFP4:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_NVFP4>
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q2_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q2_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q3_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q3_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q4_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q5_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q5_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q6_K>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XXS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_XS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ2_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ2_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ3_XXS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_XXS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ1_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ1_M:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ1_M>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ4_NL:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_NL>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ4_XS:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ4_XS>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         case GGML_TYPE_IQ3_S:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_IQ3_S>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
                  nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
-                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, stream);
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
             break;
         default:
             GGML_ABORT("fatal error");
@@ -622,7 +1140,7 @@ void ggml_cuda_mul_mat_vec_q(
     GGML_ASSERT(        nb0        == ts_dst);
     GGML_ASSERT(!ids || ids->nb[0] == ggml_type_size(ids->type));
 
-    GGML_ASSERT(!ids || ne12 == 1); // Implementation is only correct for batch size 1.
+    GGML_ASSERT(!ids || ne12 <= MMVQ_MAX_BATCH_SIZE);
 
     const float   * src1_d =       (const float   *) src1->data;
     const int32_t *  ids_d = ids ? (const int32_t *)  ids->data : nullptr;
@@ -693,11 +1211,13 @@ void ggml_cuda_mul_mat_vec_q(
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
+    const int64_t ids_stride = ids ? ids->nb[1] / ggml_type_size(ids->type) : 0;
+
     mul_mat_vec_q_switch_type(
         src0->data, src0->type, src1_q8_1.get(), ids_d, fusion_local, dst_d, ne00,
         ne01,              ncols_dst,     s01, stride_col_y,     stride_col_dst,
         ne02, nchannels_y, nchannels_dst, s02, stride_channel_y, stride_channel_dst,
-        ne03,              ne3,           s03, s13,              s3,               stream);
+        ne03,              ne3,           s03, s13,              s3,               ids_stride, stream);
 }
 
 void ggml_cuda_op_mul_mat_vec_q(
@@ -726,7 +1246,7 @@ void ggml_cuda_op_mul_mat_vec_q(
     ggml_cuda_mm_fusion_args_device fusion_local{};
     mul_mat_vec_q_switch_type(
         src0_dd_i, src0->type, src1_ddq_i, nullptr, fusion_local, dst_dd_i, ne00, row_diff, src1_ncols, stride_row_x, stride_col_y, nrows_dst,
-        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, stream);
+        1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, stream);
 
     GGML_UNUSED_VARS(src1, dst, src1_ddf_i, src1_ncols, src1_padded_row_size);
 }