perf(dflash): add opt-in exact prefill bands

server/docs/DS4.md

@@ -251,6 +251,7 @@ The runtime logs the chosen split with a `[deepseek4-split] auto-split:` banner.
 |----------|---------|
 | `DFLASH_DS4_CUDA_LAYERS` | Override the auto-split heuristic and pin the first `N` DeepSeek4 layers to CUDA. The remaining `43 - N` layers run on the Halo shard. |
 | `DFLASH_DS4_TIMING` | Enable DS4 timing logs for the layer-split parent and target-shard daemon. Useful for profiling prefill/decode breakdowns; leave unset for normal runs. |
+| `DFLASH_DS4_EXACT_PREFILL_BANDS` | Opt in to compressor-safe exact prefill bands up to four tokens on supported layer-range paths. Exact attention remains tokenwise. Leave unset for the default single-token path; `--chunk 1` is the hard fallback. |
 | `DFLASH_DS4_ROCTX` | HIP-only, default-off semantic ROCTX ranges for an external rocprof trace. The library is loaded dynamically only when set to `1`, `true`, `yes`, or `on`. |
 | `DFLASH_DS4_SPEC` / `DFLASH_DS4_DRAFT` | Enable DSpark and select its GGUF. |
 | `DFLASH_DS4_DRAFT_BACKEND` / `DFLASH_DS4_DRAFT_GPU` | Backend and device for the in-process drafter. |

server/docs/ENVIRONMENT.md

@@ -100,6 +100,7 @@ consolidation of this list into CLI flags is tracked as follow-up work.
 - `DFLASH_DS4_DRAFT_BACKEND` - deepseek4_backend.cpp
 - `DFLASH_DS4_DRAFT_GPU` - deepseek4_backend.cpp
 - `DFLASH_DS4_DSPARK_DEBUG` - deepseek4_graph.cpp
+- `DFLASH_DS4_EXACT_PREFILL_BANDS` - deepseek4_backend.cpp
 - `DFLASH_DS4_FUSED_VERIFY` - deepseek4_dspark_spec.cpp, deepseek4_loader.cpp
 - `DFLASH_DS4_HOTNESS_CSV` - deepseek4_backend.cpp
 - `DFLASH_DS4_MOE_TP` - deepseek4_backend.cpp

server/hip_compat/cuda_runtime.h

@@ -29,6 +29,7 @@ using cudaDeviceProp        = hipDeviceProp_t;
 // Error codes
 #define cudaSuccess                 hipSuccess
 #define cudaErrorInvalidValue       hipErrorInvalidValue
+#define cudaErrorNoDevice           hipErrorNoDevice
 #define cudaErrorIllegalAddress     hipErrorIllegalAddress
 #define cudaErrorAssert             hipErrorAssert
 #define cudaErrorLaunchFailure      hipErrorLaunchFailure

server/src/deepseek4/deepseek4_backend.h

@@ -33,6 +33,8 @@ int deepseek4_hybrid_prefill_chunk_tokens(
     int context_end,
     int current_cap = 0);
 
+bool deepseek4_prefill_allows_decode_graph_reuse(bool save_snapshot);
+
 class DeepSeek4Backend : public ModelBackend {
 public:
     explicit DeepSeek4Backend(const DeepSeek4BackendConfig & cfg);

server/src/deepseek4/deepseek4_fused_verify.inc

@@ -1024,7 +1024,7 @@ static int ds4_try_fused_verify_step(
         const float * embed,
         int n_tokens,
         int kv_start,
-        std::vector<float> & out_logits,
+        std::vector<float> * out_logits,
         const int32_t * token_ids,
         Ds4VerifyHooks * hooks,
         DeepSeek4StepTelemetry * telemetry,
@@ -1342,21 +1342,23 @@ static int ds4_try_fused_verify_step(
         ggml_backend_tensor_get(fg->logits, hooks->all_logits_out->data(), 0,
                                 sizeof(float) * (size_t) w.n_vocab * q);
     }
-    if (!argmax_only) {
-        out_logits.resize((size_t) w.n_vocab);
-        ggml_backend_tensor_get(fg->logits, out_logits.data(),
+    if (!argmax_only && out_logits) {
+        out_logits->resize((size_t) w.n_vocab);
+        ggml_backend_tensor_get(fg->logits, out_logits->data(),
                                 (size_t) (q - 1) * (size_t) w.n_vocab * sizeof(float),
                                 sizeof(float) * (size_t) w.n_vocab);
-    } else {
-        out_logits.clear();
+    } else if (out_logits) {
+        out_logits->clear();
     }
     if (hooks->capture_out && ex->capture && ncap > 0) {
         hooks->capture_out->resize((size_t) ncap * w.n_embd * q);
         ggml_backend_tensor_get(
             ex->capture, hooks->capture_out->data(), 0,
             sizeof(float) * hooks->capture_out->size());
     }
-    if (telemetry) {
+    if (telemetry &&
+        (argmax_only || hooks->all_logits_out || out_logits ||
+         (hooks->capture_out && ex->capture && ncap > 0))) {
         telemetry->full_graph_read_us += ds4_elapsed_us(read_t0, Ds4TimingClock::now());
     }
     return 1;