Detect operating environment and adjust certain behaviors.

This commit introduces a new environment detection mechanism based on
process name and environment variables.

It uses this information (initially) for the purposes of tuning
concurrent model behaviors on cloud.

Signed-off-by: Jacob Howard <jacob.howard@docker.com>

Author: xenoscopic

pkg/environment/environment.go

@@ -0,0 +1,61 @@
+package environment
+
+import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"sync"
+)
+
+// Environment encodes the operating environment for the model runner.
+type Environment uint8
+
+const (
+	// EnvironmentUnknown represents an unknown environment in which basic,
+	// non-specialized defaults should be used.
+	EnvironmentUnknown Environment = iota
+	// EnvironmentDesktop represents a Docker Desktop environment.
+	EnvironmentDesktop
+	// EnvironmentMoby represents a Moby engine environment, if installed via
+	// the model CLI.
+	EnvironmentMoby
+	// EnvironmentCloud represents a Docker Cloud environment, if installed via
+	// the model CLI.
+	EnvironmentCloud
+)
+
+// environment is the cached environment.
+var environment Environment
+
+// environmentOnce guards initialization of environment.
+var environmentOnce sync.Once
+
+// isDockerBackend checks if an executable path is com.docker.backend.
+func isDockerBackend(path string) bool {
+	_, leaf := filepath.Split(path)
+	if runtime.GOOS == "windows" {
+		return leaf == "com.docker.backend.exe"
+	}
+	return leaf == "com.docker.backend"
+}
+
+// Get returns the current environment type.
+func Get() Environment {
+	environmentOnce.Do(func() {
+		// Check if we're running in a Docker Desktop backend process.
+		if executable, err := os.Executable(); err == nil && isDockerBackend(executable) {
+			environment = EnvironmentDesktop
+			return
+		}
+
+		// Look for a MODEL_RUNNER_ENVIRONMENT variable. If none is set or it's
+		// invalid, then leave the environment unknown.
+		switch os.Getenv("MODEL_RUNNER_ENVIRONMENT") {
+		case "moby":
+			environment = EnvironmentMoby
+		case "cloud":
+			environment = EnvironmentCloud
+		}
+	})
+	return environment
+}

pkg/inference/scheduling/loader.go

@@ -7,6 +7,7 @@ import (
 	"runtime"
 	"time"
 
+	"github.com/docker/model-runner/pkg/environment"
 	"github.com/docker/model-runner/pkg/inference"
 	"github.com/docker/model-runner/pkg/inference/models"
 	"github.com/docker/model-runner/pkg/logging"
@@ -110,7 +111,12 @@ func newLoader(
 	// VRAM size here (and potentially even reserving a portion of it) and
 	// computing model size through estimation (using parameter count and
 	// quantization data type size).
+	//
+	// HACK: On GPU-enabled cloud engines, we'll temporarily bump this to 2.
 	totalMemory := uint64(1)
+	if environment.Get() == EnvironmentCloud && os.Getenv("NVIDIA_VISIBLE_DEVICES") != "" {
+		totalMemory = 2
+	}
 
 	// Create the loader.
 	l := &loader{