feat(deepseek4): run the vision encoder on a second GPU and stream images into prefill

docs/image-input.md

@@ -66,6 +66,25 @@ launch plus `--mmproj`: the Vision file replaces
 `DeepSeek-V4-Flash-0731-ROCMFPX-MIX-STRIX.gguf` for text as well and decodes
 at least as fast (numbers below). For R9700 + Strix Halo see [DS4V](#ds4v) below.
 
+With an R9700 in the same box, run the image encoder there while the model
+stays on the Strix Halo: expose both GPUs, point `--target-device` at the Strix
+Halo and add `--mmproj-device` with the R9700 (on lucebox6, without
+`HIP_VISIBLE_DEVICES`, that is `--target-device hip:1 --mmproj-device hip:0`).
+The encoder then runs about twice as fast and streams each image into prefill
+as soon as it is encoded, so the Strix Halo never waits for the next one:
+
+| Images | Prompt tokens | Encoder on the Strix Halo | Encoder on the R9700 |
+| --- | --- | --- | --- |
+| 1 | 126 | 2.97 s | 2.94 s |
+| 4 | 942 | 11.2 s | 9.1 s |
+| 8 | 2,262 | 27.7 s | 19.2 s |
+| 16 | 4,358 | 51.8 s | 34.6 s |
+
+Time to the first token, with the published launch above and ChartQA charts.
+Answers are identical in both layouts. `--mmproj-device` applies to this one-GPU
+layout; with the experts split across both GPUs the encoder already runs on the
+R9700.
+
 ### Send an image
 
 ```bash
@@ -101,7 +120,7 @@ Use `POST /v1/chat/completions` with user-message content parts in display order
 
 Only base64 JPEG/PNG data URLs are supported. Remote URLs, images outside user
 content arrays, and image parts through other API formats are rejected. A
-request carries at most four images, 16 MiB encoded each and 32 MiB combined.
+request carries at most 16 images, 16 MiB encoded each and 32 MiB combined.
 Decoder pixel and aspect limits also apply. A model's image marker cannot be supplied
 as ordinary text.
 

server/src/common/backend_args.h

@@ -56,6 +56,9 @@ struct BackendArgs {
 
     // Optional: vision projector .gguf (deepseek4 only)
     std::optional<std::string> mmproj_path;
+    // Optional: GPU for the vision encoder when it should not share the
+    // target's (deepseek4, one-GPU layout only).
+    std::optional<DevicePlacement> mmproj_device;
 
     // Device placement
     DevicePlacement device;

server/src/common/backend_factory.cpp

@@ -370,6 +370,7 @@ std::unique_ptr<ModelBackend> construct_backend(
             DeepSeek4BackendConfig cfg;
             cfg.model_path = model.path;
             cfg.mmproj_path = model.mmproj_path.value_or("");
+            cfg.mmproj_gpu  = model.mmproj_device ? model.mmproj_device->gpu : -1;
             cfg.device     = placement.target;
             cfg.stream_fd  = execution.stream_fd;
             cfg.max_ctx    = placement.target.max_ctx;

server/src/common/backend_factory.h

@@ -37,6 +37,7 @@ class BackendPlan final {
     struct Model {
         std::string path;
         std::optional<std::string> mmproj_path;
+        std::optional<DevicePlacement> mmproj_device;
         GgufModelInfo metadata;
     };
 

server/src/common/backend_plan.cpp

@@ -131,6 +131,7 @@ BackendPreparation BackendPlanBuilder::resolve(
     BackendPlan plan;
     plan.model_.path = std::move(args.model_path);
     plan.model_.mmproj_path = std::move(args.mmproj_path);
+    plan.model_.mmproj_device = std::move(args.mmproj_device);
     plan.model_.metadata = std::move(model);
 
     plan.placement_.target = std::move(args.device);

server/src/common/feature_gate.cpp

@@ -59,6 +59,14 @@ std::string check_feature_compatibility(
             return "--mmproj with DeepSeek4 requires a HIP backend";
         }
     }
+    if (args.mmproj_device.has_value()) {
+        if (!args.mmproj_path.has_value() || arch != "deepseek4" ||
+            args.mmproj_device->backend != PlacementBackend::Hip ||
+            args.mmproj_device->gpu == args.device.gpu) {
+            return "--mmproj-device needs --mmproj, a DeepSeek4 target, and a HIP GPU "
+                   "other than the target's";
+        }
+    }
 
     // ── PFlash enablement × drafter model
     if (admission.pflash_enabled &&

server/src/common/image_prompt.h

@@ -1,12 +1,17 @@
 #pragma once
 
+#include <cstddef>
 #include <cstdint>
 #include <memory>
 #include <string>
 #include <vector>
 
 namespace luce::common {
 
+// Most images one request may carry, for every vision backend and the HTTP
+// transport. Each image still has its own token and byte bounds.
+inline constexpr size_t MAX_REQUEST_IMAGES = 16;
+
 struct EncodedImage {
     std::string mime_type;
     std::vector<uint8_t> bytes;