From 1df37b654b3c36ff4ac40da0dea00700b3ae803f Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Fri, 22 Mar 2024 19:46:35 +0200
Subject: [PATCH] iq1_m: CUDA dequantize works

Very 1st shot I get PPL = 9.76 for LLaMA-v2-7B.
---
 ggml-cuda.cu | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/ggml-cuda.cu b/ggml-cuda.cu
index 4f50c9f9f..48232b6e1 100644
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -615,6 +615,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -643,6 +644,7 @@ static int64_t get_row_rounding(ggml_type type, const std::array<float, GGML_CUD
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_IQ3_XXS:
         case GGML_TYPE_IQ1_S:
+        case GGML_TYPE_IQ1_M:
         case GGML_TYPE_IQ4_NL:
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
@@ -2560,7 +2562,7 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
                 ggml_type a_type = a->type;
                 if (a_type == GGML_TYPE_IQ2_XXS || a_type == GGML_TYPE_IQ2_XS || a_type == GGML_TYPE_IQ3_XXS ||
                     a_type == GGML_TYPE_IQ1_S   || a_type == GGML_TYPE_IQ4_NL || a_type == GGML_TYPE_IQ3_S   ||
-                    a_type == GGML_TYPE_IQ2_S   || a_type == GGML_TYPE_IQ4_XS) {
+                    a_type == GGML_TYPE_IQ1_M   || a_type == GGML_TYPE_IQ2_S  || a_type == GGML_TYPE_IQ4_XS) {
                     if (b->ne[1] == 1 && ggml_nrows(b) > 1) {
                         return false;
                     }