feat(ggml): vendor ROCmFPX quant and DS4 ops

.github/workflows/ci.yml

@@ -273,3 +273,19 @@ jobs:
           /opt/rocm/bin/hipcc --offload-arch=gfx1151 -O2 \
             -o "$RUNNER_TEMP/hip_smoke" .github/ci/hip_smoke.cpp
           "$RUNNER_TEMP/hip_smoke"
+
+      - name: Build + test ROCmFP formats
+        run: |
+          cmake -S server -B "$RUNNER_TEMP/rocmfp-build" \
+            -DDFLASH27B_GPU_BACKEND=hip \
+            -DDFLASH27B_HIP_ARCHITECTURES=gfx1151 \
+            -DDFLASH27B_SERVER=OFF \
+            -DDFLASH27B_TESTS=ON \
+            -DCMAKE_BUILD_TYPE=Release \
+            -DCMAKE_HIP_FLAGS=-DDFLASH_WAVE_SIZE=32
+          cmake --build "$RUNNER_TEMP/rocmfp-build" \
+            --target test_rocmfp4 test_rocmfpx test_rocmfp4_hip_tail \
+            --parallel 8
+          ctest --test-dir "$RUNNER_TEMP/rocmfp-build" \
+            --output-on-failure \
+            -R 'rocmfp4_reference|rocmfpx_reference|rocmfp4_hip_tail'

server/CMakeLists.txt

@@ -655,6 +655,36 @@ if(DFLASH27B_TESTS)
             ${DFLASH27B_GGML_BACKEND_TARGET} ggml ggml-base)
         add_test(NAME cuda_pool_shutdown COMMAND test_cuda_pool_shutdown)
     endif()
+
+    add_executable(test_rocmfp4 deps/llama.cpp/ggml/rocmfp4/test_rocmfp4.c)
+    target_link_libraries(test_rocmfp4 PRIVATE ggml-base)
+    if(UNIX)
+        target_link_libraries(test_rocmfp4 PRIVATE m)
+    endif()
+    add_test(NAME rocmfp4_reference COMMAND test_rocmfp4)
+
+    add_executable(test_rocmfpx deps/llama.cpp/ggml/rocmfpx/test_rocmfpx.c)
+    target_link_libraries(test_rocmfpx PRIVATE ggml-base)
+    if(UNIX)
+        target_link_libraries(test_rocmfpx PRIVATE m)
+    endif()
+    add_test(NAME rocmfpx_reference COMMAND test_rocmfpx)
+
+    if(DFLASH27B_GPU_BACKEND STREQUAL "hip")
+        add_executable(test_rocmfp4_hip_tail test/test_rocmfp4_hip_tail.cpp)
+        set_source_files_properties(test/test_rocmfp4_hip_tail.cpp PROPERTIES LANGUAGE HIP)
+        set_target_properties(test_rocmfp4_hip_tail PROPERTIES HIP_ARCHITECTURES "${_dflash_archs}")
+        target_include_directories(test_rocmfp4_hip_tail PRIVATE
+            ${CMAKE_CURRENT_SOURCE_DIR}/deps/llama.cpp/ggml/include
+            ${CMAKE_CURRENT_SOURCE_DIR}/deps/llama.cpp/ggml/rocmfp4)
+        target_link_libraries(test_rocmfp4_hip_tail PRIVATE
+            ggml
+            ggml-base
+            ${DFLASH27B_GGML_BACKEND_TARGET}
+            hip::host)
+        add_test(NAME rocmfp4_hip_tail COMMAND test_rocmfp4_hip_tail)
+    endif()
+
     if(DFLASH27B_GPU_BACKEND STREQUAL "cuda" AND _dflash_cuda_min_sm GREATER_EQUAL 80 AND EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/test/test_flashprefill_kernels.cpp")
         add_executable(test_flashprefill_kernels test/test_flashprefill_kernels.cpp)
         set_target_properties(test_flashprefill_kernels PROPERTIES CUDA_ARCHITECTURES "${_dflash_archs}")

server/deps/llama.cpp/ggml/include/ggml.h

@@ -430,7 +430,13 @@ extern "C" {
         GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
         GGML_TYPE_Q1_0    = 41,
         GGML_TYPE_TQ3_0   = 42,  // TurboQuant 3.5 bpv (3-bit Lloyd-Max + FWHT rotation)
-        GGML_TYPE_COUNT   = 43,
+        GGML_TYPE_Q4_0_ROCMFP4      = 100,
+        GGML_TYPE_Q4_0_ROCMFP4_FAST = 101,
+        GGML_TYPE_Q6_0_ROCMFPX      = 102,
+        GGML_TYPE_Q8_0_ROCMFPX      = 103,
+        GGML_TYPE_Q3_0_ROCMFPX      = 104,
+        GGML_TYPE_Q2_0_ROCMFP2      = 107,
+        GGML_TYPE_COUNT   = 108,
     };
 
     // precision
@@ -468,6 +474,18 @@ extern "C" {
         GGML_FTYPE_MOSTLY_MXFP4   = 25, // except 1d tensors
         GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q1_0    = 27, // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4          = 100,
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4_LEAN     = 101,
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4_COHERENT = 102,
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4_FAST     = 103,
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4_FAST_COHERENT = 104,
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4_STRIX    = 105,
+        GGML_FTYPE_MOSTLY_Q4_0_ROCMFP4_STRIX_LEAN = 106,
+        GGML_FTYPE_MOSTLY_Q6_0_ROCMFPX          = 110,
+        GGML_FTYPE_MOSTLY_Q8_0_ROCMFPX          = 111,
+        GGML_FTYPE_MOSTLY_Q3_0_ROCMFPX          = 112,
+        GGML_FTYPE_MOSTLY_Q2_0_ROCMFP2          = 118,
+        GGML_FTYPE_MOSTLY_Q2_0_ROCMFP2_STRIX    = 119,
     };
 
     // available tensor operations:
@@ -583,6 +601,8 @@ extern "C" {
 
         GGML_OP_MOE_FUSED,  // Fused MoE FFN: gate+up+swiglu+down+weighted_sum+shared_expert
 
+        GGML_OP_DS4_HC,  // Fused DeepSeek4 hyper-connection pre/post/out mixing
+
         GGML_OP_COUNT,
     };
 
@@ -618,6 +638,7 @@ extern "C" {
         GGML_GLU_OP_GEGLU,
         GGML_GLU_OP_SWIGLU,
         GGML_GLU_OP_SWIGLU_OAI,
+        GGML_GLU_OP_SWIGLU_DS4,
         GGML_GLU_OP_GEGLU_ERF,
         GGML_GLU_OP_GEGLU_QUICK,
 
@@ -1339,6 +1360,12 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_swiglu_ds4_split(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * gate,
+            struct ggml_tensor  * up,
+            float                 clamp);
+
     GGML_API struct ggml_tensor * ggml_geglu_erf_split(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2399,6 +2426,37 @@ extern "C" {
             struct ggml_tensor  * experts,
             struct ggml_tensor  * expert_weights);
 
+    // Fused DeepSeek4 hyper-connection helpers (decode, n_tokens == 1).
+    // ggml_ds4_hc_pre: mix[2*n_hc+n_hc^2] + base + hc_state[n_embd*n_hc] ->
+    //   dst[n_embd + 2*n_hc + n_hc^2] = { working, split(pre,post,comb) }
+    GGML_API struct ggml_tensor * ggml_ds4_hc_pre(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * mix,
+            struct ggml_tensor  * base,
+            struct ggml_tensor  * hc_state,
+            int                   n_hc,
+            int                   sinkhorn_iters,
+            float                 pre_scale,
+            float                 post_scale,
+            float                 comb_scale);
+
+    // ggml_ds4_hc_post: residual hc_state + block_out + split -> new hc_state
+    GGML_API struct ggml_tensor * ggml_ds4_hc_post(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * residual_hc,
+            struct ggml_tensor  * block_out,
+            struct ggml_tensor  * split,
+            int                   n_hc);
+
+    // ggml_ds4_hc_out: output-stage merge of hc streams into one embedding
+    GGML_API struct ggml_tensor * ggml_ds4_hc_out(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * mix,
+            struct ggml_tensor  * base,
+            struct ggml_tensor  * hc_state,
+            int                   n_hc,
+            float                 pre_scale);
+
     // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,

server/deps/llama.cpp/ggml/rocmfp4/README.md

@@ -0,0 +1,40 @@
+# ROCmFP4
+
+This directory implements two GGUF weight formats used by the ROCm/HIP
+backend:
+
+| GGML type | Block layout | Block size | Bits/weight |
+|---|---|---:|---:|
+| `Q4_0_ROCMFP4` | 32 packed 4-bit codes and one UE4M3 scale per 16 weights | 18 bytes | 4.50 |
+| `Q4_0_ROCMFP4_FAST` | 32 packed 4-bit codes and one UE4M3 scale per 32 weights | 17 bytes | 4.25 |
+
+Both layouts use the signed Codebook10 levels
+`0, +/-1, +/-2, +/-3, +/-4, +/-6, +/-8, +/-10`. Scale bytes are finite
+unsigned UE4M3 values; validation rejects `0x7f` and values with the sign bit
+set.
+
+## Implementation
+
+- `rocmfp4.c` contains deterministic CPU reference quantization,
+  dequantization, validation, and vector-dot functions.
+- `rocmfp4_hip_codebook.cuh` and `rocmfp4_hip_scale.cuh` contain device
+  helpers shared by the ggml HIP kernels.
+- Backend dispatch is integrated in `ggml-cuda`, which is also the source tree
+  used by ggml's HIP build.
+
+The quantizer selects UE4M3 scales by minimizing reconstruction error. When an
+importance matrix is supplied, finite positive weights are used in the error
+metric. Non-finite source values quantize to zero and non-finite or non-positive
+importance weights are ignored.
+
+Vulkan kernels are not implemented in this vendored build.
+
+## Tests
+
+With `DFLASH27B_TESTS=ON`:
+
+- `test_rocmfp4` checks CPU reference round trips, validation, and imatrix
+  handling.
+- `test_rocmfp4_hip_tail` compares HIP conversion with the CPU reference
+  byte-for-byte and checks output bounds for row sizes that are and are not
+  multiples of 256.