llm: introduce k/v context quantization (vRAM improvements) (#6279)

llm/server.go

@@ -214,15 +214,36 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 		params = append(params, "--threads", strconv.Itoa(defaultThreads))
 	}
 
-	flashAttnEnabled := envconfig.FlashAttention()
+	fa := envconfig.FlashAttention()
+	if fa && !gpus.FlashAttentionSupported() {
+		slog.Warn("flash attention enabled but not supported by gpu")
+		fa = false
+	}
 
-	for _, g := range gpus {
-		// only cuda (compute capability 7+) and metal support flash attention
-		if g.Library != "metal" && (g.Library != "cuda" || g.DriverMajor < 7) {
-			flashAttnEnabled = false
+	if fa && !ggml.SupportsFlashAttention() {
+		slog.Warn("flash attention enabled but not supported by model")
+		fa = false
+	}
+
+	kvct := envconfig.KvCacheType()
+
+	if fa {
+		slog.Info("enabling flash attention")
+		params = append(params, "--flash-attn")
+
+		// Flash Attention also supports kv cache quantization
+		// Enable if the requested and kv cache type is supported by the model
+		if kvct != "" && ggml.SupportsKVCacheType(kvct) {
+			params = append(params, "--kv-cache-type", kvct)
+		} else {
+			slog.Warn("kv cache type not supported by model", "type", kvct)
 		}
+	} else if kvct != "" && kvct != "f16" {
+		slog.Warn("quantized kv cache requested but flash attention disabled", "type", kvct)
+	}
 
-		// mmap has issues with partial offloading on metal
+	// mmap has issues with partial offloading on metal
+	for _, g := range gpus {
 		if g.Library == "metal" &&
 			uint64(opts.NumGPU) > 0 &&
 			uint64(opts.NumGPU) < ggml.KV().BlockCount()+1 {
@@ -231,10 +252,6 @@ func NewLlamaServer(gpus discover.GpuInfoList, model string, ggml *GGML, adapter
 		}
 	}
 
-	if flashAttnEnabled {
-		params = append(params, "--flash-attn")
-	}
-
 	// Windows CUDA should not use mmap for best performance
 	// Linux  with a model larger than free space, mmap leads to thrashing
 	// For CPU loads we want the memory to be allocated, not FS cache