feat: phased patch system for Intel Arc GPU performance fixes

3-model council (GLM-5.1, Minimax-M2.7, Kimi k2p5) analyzed Intel Arc GPU
performance issues and produced patches for llama.cpp:

Phase 1 - SYCL Sync: Enable graph execution by default (GGML_SYCL_DISABLE_GRAPH)
Phase 2 - SYCL Kernel: Fix VER_GEN12/13 thresholds, tune DMMV_X/MMV_Y
Phase 3 - Vulkan Intel: Arc 140T device-ID Xe2 override

Includes:
- Phased apply script (apply-phase.sh [1|2|3|all])
- Master apply script with --status/--reverse/--dry-run
- Per-phase READMEs with testing checklists
- Council deliberation logs (gitignored in logs/)

Verified: all patches apply/reverse cleanly via git apply.
Static verification: VER_GEN arithmetic and DMMV_X divisibility pass.

.gitignore

@@ -1 +1,2 @@
 .DS_Store
+logs/

repos/patch/README.md

@@ -0,0 +1,93 @@
+# Intel GPU Diagnosis Patches
+
+Patches generated by a 3-model council (GLM-5.1, Minimax-M2.7, Kimi k2p5) analyzing
+Intel Arc GPU performance issues in llama.cpp. Non-overlapping training data ensures
+blind spots are caught through cross-review.
+
+## Quick Start
+
+```bash
+# Apply all phases at once:
+cd repos
+./patch/apply.sh
+
+# Apply one phase at a time (recommended for testing):
+./patch/apply-phase.sh 1              # Apply phase 1
+./patch/apply-phase.sh 2              # Apply phase 2 (after testing phase 1)
+./patch/apply-phase.sh 3              # Apply phase 3 (after testing phase 2)
+
+# Dry-run / reverse:
+./patch/apply-phase.sh 1 --dry-run
+./patch/apply-phase.sh 2 --reverse
+
+# Check what's applied:
+./patch/apply.sh --status
+```
+
+## Phases
+
+### Phase 1 — SYCL Sync (safest, highest impact)
+| Patch | File | Change | Decision |
+|-------|------|--------|----------|
+| 0001 | ggml-sycl.cpp:217 | Graph default 1→0 | Approved 3/3 |
+
+Enables SYCL graph execution by default. Eliminates 8 blocking `.wait()` calls.
+Expected 10-30% token generation speedup for single-GPU dense LLMs.
+
+### Phase 2 — SYCL Kernel Tuning (depends on Phase 1)
+| Patch | File | Change | Decision |
+|-------|------|--------|----------|
+| 0001 | common.hpp:90-91 | VER_GEN12 1M→1200, VER_GEN13→1300 | Approved 3/3 |
+| 0002 | presets.hpp:57,60 | DMMV_X 32→64, MMV_Y 1→2 | Approved 3/3, needs bench |
+| 0003 | common.hpp:106,109 | DMMV_X 32→64, MMV_Y 1→2 | Approved 3/3, needs bench |
+
+Fixes the VER_GEN12 placeholder (1M) that routed all Intel GPUs to NVIDIA Ampere paths.
+Tunes DMMV thread parameters for Arc hardware. Expected 5-15% additional improvement.
+
+### Phase 3 — Vulkan Intel (independent of Phase 1/2)
+| Patch | File | Change | Decision |
+|-------|------|--------|----------|
+| 0001 | ggml-vulkan.cpp:302,349 | Arc 140T device-ID Xe2 override | Approved 1/3 |
+
+Fixes Arrow Lake H misdetection as non-Xe2. Enables cooperative matrix.
+Only affects Arc 140T systems.
+
+## Testing Protocol
+
+On Intel Arc GPU test machine:
+
+```bash
+cd repos
+
+# Apply one phase
+./patch/apply-phase.sh 1
+
+# Build
+source /opt/intel/oneapi/setvars.sh
+cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
+  -DCMAKE_BUILD_TYPE=Release -DGGML_AVX2=ON -DGGML_AVX=ON -DGGML_FMA=ON -DGGML_F16C=ON
+cmake --build build -j$(nproc)
+
+# Test
+ctest --test-dir build --output-on-failure
+
+# Benchmark before/after each phase
+./build/bin/llama-bench -m <model> -ngl 99 -p 512 -n 128
+```
+
+## Council Deliberations
+
+Stored in `../logs/` (gitignored). Key files:
+- `logs/decisions.md` — All 6 council decisions with rationale
+- `logs/M-sync-overhead-20260415.md` — Agent-M sync analysis
+- `logs/K-kernel-tuning-20260415.md` — Agent-K kernel tuning analysis
+- `logs/M-review-K-20260415.md` — Cross-review
+- `logs/K-review-M-20260415.md` — Cross-review (caught memset_tensor error)
+
+## Deferred to Future Phases
+
+- Q4_K DMMV reorder (medium complexity)
+- Q6_K DMMV reorder (medium complexity)
+- Q5_K reorder for both DMMV and MMVQ (hard)
+- Host-buffer double-copy elimination
+- Async memory ops decoupled from graph

repos/patch/apply-phase.sh

@@ -0,0 +1,147 @@
+#!/usr/bin/env bash
+#
+# Apply patches for a single phase.
+# Usage:
+#   apply-phase.sh 1              # Apply phase 1 patches
+#   apply-phase.sh 2 --dry-run    # Dry-run phase 2
+#   apply-phase.sh 2 --reverse    # Reverse phase 2
+#   apply-phase.sh all            # Apply all phases in order
+#
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPOS_DIR="$(dirname "$SCRIPT_DIR")"
+
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[0;33m'
+BLUE='\033[0;34m'
+NC='\033[0m'
+
+PHASE_NAME=""
+DRY_RUN=0
+REVERSE=0
+
+if [[ $# -lt 1 ]]; then
+    echo "Usage: $0 <phase-number|all> [--dry-run|--reverse]"
+    echo "Phases: 1=sycl-sync, 2=sycl-kernel, 3=vulkan-intel"
+    exit 1
+fi
+
+PHASE_ARG="$1"
+shift
+
+for arg in "$@"; do
+    case "$arg" in
+        --dry-run) DRY_RUN=1 ;;
+        --reverse) REVERSE=1 ;;
+    esac
+done
+
+apply_phase() {
+    local phase_num="$1"
+    local phase_dir=""
+
+    case "$phase_num" in
+        1) phase_dir="phase1-sycl-sync"; PHASE_NAME="SYCL Sync (graph + async)" ;;
+        2) phase_dir="phase2-sycl-kernel"; PHASE_NAME="SYCL Kernel (VER_GEN + DMMV tuning)" ;;
+        3) phase_dir="phase3-vulkan-intel"; PHASE_NAME="Vulkan Intel (Arc 140T Xe2 override)" ;;
+        *)
+            echo -e "${RED}Unknown phase: $phase_num${NC}"
+            exit 1
+            ;;
+    esac
+
+    local patch_dir="$SCRIPT_DIR/$phase_dir"
+    local target_repo="$REPOS_DIR/llama.cpp"
+
+    if [[ ! -d "$target_repo" ]]; then
+        echo -e "${RED}Repo not found: $target_repo${NC}"
+        exit 1
+    fi
+
+    if [[ ! -d "$patch_dir" ]]; then
+        echo -e "${YELLOW}No patch directory: $patch_dir${NC}"
+        return 0
+    fi
+
+    local patches=($(ls "$patch_dir"/*.patch 2>/dev/null | sort))
+    if [[ ${#patches[@]} -eq 0 ]]; then
+        echo -e "${YELLOW}No patches in $patch_dir${NC}"
+        return 0
+    fi
+
+    local action="APPLYING"
+    local git_flag=""
+    if [[ $DRY_RUN -eq 1 ]]; then
+        action="CHECKING"
+        git_flag="--check"
+    fi
+    if [[ $REVERSE -eq 1 ]]; then
+        action="REVERSING"
+        git_flag="-R"
+        if [[ $DRY_RUN -eq 1 ]]; then
+            git_flag="-R --check"
+        fi
+    fi
+
+    local applied=0 failed=0
+
+    for patch_file in "${patches[@]}"; do
+        local patch_name=$(basename "$patch_file")
+        echo -n -e "  ${BLUE}${action}${NC} ${phase_dir}/${patch_name} ... "
+
+        if (cd "$target_repo" && git apply $git_flag "$patch_file" 2>/dev/null); then
+            if [[ $DRY_RUN -eq 1 ]]; then
+                echo -e "${GREEN}[OK dry-run]${NC}"
+            else
+                echo -e "${GREEN}[OK]${NC}"
+            fi
+            ((applied++)) || true
+        else
+            if [[ $DRY_RUN -eq 1 ]]; then
+                echo -e "${RED}[FAIL dry-run]${NC}"
+            else
+                echo -e "${RED}[FAIL]${NC}"
+            fi
+            ((failed++)) || true
+        fi
+    done
+
+    echo -e "  ${GREEN}$applied ok${NC}, ${RED}$failed failed${NC}"
+    return $failed
+}
+
+if [[ "$PHASE_ARG" == "all" ]]; then
+    echo -e "${BLUE}=== All Phases ===${NC}"
+    if [[ $DRY_RUN -eq 1 ]]; then echo -e "(dry-run)"; fi
+    if [[ $REVERSE -eq 1 ]]; then echo -e "(reversing in reverse order: 3, 2, 1)"; fi
+    echo ""
+    total_fail=0
+
+    if [[ $REVERSE -eq 1 ]]; then
+        for p in 3 2 1; do
+            echo -e "${BLUE}[Phase $p]${NC}"
+            apply_phase $p || ((total_fail++)) || true
+            echo ""
+        done
+    else
+        for p in 1 2 3; do
+            echo -e "${BLUE}[Phase $p]${NC}"
+            apply_phase $p || ((total_fail++)) || true
+            echo ""
+        done
+    fi
+
+    if [[ $total_fail -gt 0 ]]; then
+        echo -e "${RED}$total_fail phase(s) had failures.${NC}"
+        exit 1
+    fi
+    echo -e "${GREEN}All phases complete.${NC}"
+else
+    echo -e "${BLUE}=== Phase $PHASE_ARG: $PHASE_NAME ===${NC}"
+    if [[ $DRY_RUN -eq 1 ]]; then echo -e "(dry-run)"; fi
+    if [[ $REVERSE -eq 1 ]]; then echo -e "(reversing)"; fi
+    echo ""
+    apply_phase "$PHASE_ARG"
+fi

repos/patch/apply.sh

@@ -0,0 +1,47 @@
+#!/usr/bin/env bash
+#
+# apply.sh — Apply all Intel GPU diagnosis patches
+#
+# Delegates to apply-phase.sh. See that script for options.
+#
+set -euo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[0;33m'
+NC='\033[0m'
+
+# Parse args to detect --reverse and pass through
+REVERSE=0
+DRY_RUN=0
+for arg in "$@"; do
+    case "$arg" in
+        --reverse) REVERSE=1 ;;
+        --dry-run) DRY_RUN=1 ;;
+        --status)
+            # Status mode — check each patch
+            echo "Patch status:"
+            for phase_dir in "$SCRIPT_DIR"/phase*/; do
+                phase_name=$(basename "$phase_dir")
+                echo ""
+                echo "  [$phase_name]"
+                for patch_file in "$phase_dir"*.patch; do
+                    [[ -f "$patch_file" ]] || continue
+                    patch_name=$(basename "$patch_file")
+                    target_repo="$(dirname "$SCRIPT_DIR")/llama.cpp"
+                    if (cd "$target_repo" && git apply --check "$patch_file" 2>/dev/null); then
+                        echo -e "    ${YELLOW}NOT APPLIED${NC} $patch_name"
+                    elif (cd "$target_repo" && git apply --check -R "$patch_file" 2>/dev/null); then
+                        echo -e "    ${GREEN}APPLIED${NC}    $patch_name"
+                    else
+                        echo -e "    ${RED}CONFLICT${NC}   $patch_name"
+                    fi
+                done
+            done
+            exit 0
+            ;;
+    esac
+done
+
+exec "$SCRIPT_DIR/apply-phase.sh" all "$@"

repos/patch/phase1-sycl-sync/0001-enable-sycl-graph-by-default.patch

@@ -0,0 +1,13 @@
+diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp
+index ea79d25..e93201c 100644
+--- a/ggml/src/ggml-sycl/ggml-sycl.cpp
++++ b/ggml/src/ggml-sycl/ggml-sycl.cpp
+@@ -214,7 +214,7 @@ static void ggml_check_sycl() try {
+     if (!initialized) {
+         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
+         g_ggml_sycl_disable_optimize = get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
+-        g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
++        g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 0);
+         g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0);
+         g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
+ 

repos/patch/phase1-sycl-sync/README.md

@@ -0,0 +1,26 @@
+# Phase 1 — SYCL Synchronization
+
+## 0001-enable-sycl-graph-by-default.patch
+
+Changes `GGML_SYCL_DISABLE_GRAPH` default from 1 (disabled) to 0 (enabled).
+
+### What it does
+- Enables SYCL graph execution for single-GPU dense LLM inference
+- Enables async memory operations (tied to graph support in upstream code)
+- Eliminates 8 blocking `.wait()` calls in reorder functions (Q4_0, Q8_0, Q4_K, Q6_K)
+
+### What it does NOT affect
+- MoE models (MUL_MAT_ID) — `check_graph_compatibility()` auto-disables graphs
+- CONCAT operations — auto-disabled
+- Multi-GPU setups — always disabled
+- Users can override: `GGML_SYCL_DISABLE_GRAPH=1`
+
+### Expected impact
+10-30% token generation speedup on single-GPU dense LLM inference.
+
+### Testing checklist
+- [ ] Build succeeds with `-DGGML_SYCL=ON`
+- [ ] `GGML_SYCL_DEBUG=1` shows "SYCL-GRAPH" messages for dense models
+- [ ] Dense model inference produces correct output
+- [ ] MoE model falls back gracefully (logs "disabling SYCL graphs")
+- [ ] `GGML_SYCL_DISABLE_GRAPH=1` restores old behavior

repos/patch/phase2-sycl-kernel/0001-fix-ver-gen-thresholds.patch

@@ -0,0 +1,15 @@
+diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp
+index fd84c91..48cc91f 100644
+--- a/ggml/src/ggml-sycl/common.hpp
++++ b/ggml/src/ggml-sycl/common.hpp
+@@ -87,8 +87,8 @@ extern int g_ggml_sycl_enable_flash_attention;
+ #define __SYCL_ARCH__ DPCT_COMPATIBILITY_TEMP
+ #define VER_4VEC 610 // todo for hardware optimize.
+ #define VER_GEN9 700 // todo for hardware optimize.
+-#define VER_GEN12 1000000 // todo for hardware optimize.
+-#define VER_GEN13 (VER_GEN12 + 1030) // todo for hardware optimize.
++#define VER_GEN12 1200 // Intel discrete GPUs (DG2/Alchemist, major=12)
++#define VER_GEN13 1300 // Intel Xe2/Battlemage GPUs (major=13)
+ 
+ #define GGML_SYCL_MAX_NODES 8192 // TODO: adapt to hardwares
+ 

repos/patch/phase2-sycl-kernel/0002-tune-dmmv-xy-for-arc.patch

@@ -0,0 +1,17 @@
+diff --git a/ggml/src/ggml-sycl/presets.hpp b/ggml/src/ggml-sycl/presets.hpp
+index dc4dad1..a2e4ce2 100644
+--- a/ggml/src/ggml-sycl/presets.hpp
++++ b/ggml/src/ggml-sycl/presets.hpp
+@@ -54,10 +54,10 @@
+ 
+ // dmmv = dequantize_mul_mat_vec
+ #ifndef GGML_SYCL_DMMV_X
+-#define GGML_SYCL_DMMV_X 32
++#define GGML_SYCL_DMMV_X 64
+ #endif
+ #ifndef GGML_SYCL_MMV_Y
+-#define GGML_SYCL_MMV_Y 1
++#define GGML_SYCL_MMV_Y 2
+ #endif
+ 
+ #ifndef K_QUANTS_PER_ITERATION

repos/patch/phase2-sycl-kernel/0003-tune-dmmv-xy-common-hpp.patch

@@ -0,0 +1,17 @@
+diff --git a/ggml/src/ggml-sycl/common.hpp b/ggml/src/ggml-sycl/common.hpp
+index fd84c91..dd5cf1a 100644
+--- a/ggml/src/ggml-sycl/common.hpp
++++ b/ggml/src/ggml-sycl/common.hpp
+@@ -103,10 +103,10 @@ extern int g_ggml_sycl_enable_flash_attention;
+ 
+ // dmmv = dequantize_mul_mat_vec
+ #ifndef GGML_SYCL_DMMV_X
+-#define GGML_SYCL_DMMV_X 32
++#define GGML_SYCL_DMMV_X 64
+ #endif
+ #ifndef GGML_SYCL_MMV_Y
+-#define GGML_SYCL_MMV_Y 1
++#define GGML_SYCL_MMV_Y 2
+ #endif
+ 
+ typedef sycl::queue *queue_ptr;

repos/patch/phase2-sycl-kernel/README.md

@@ -0,0 +1,37 @@
+# Phase 2 — SYCL Kernel Tuning
+
+**Depends on:** Phase 1 (should be applied and tested first)
+
+## 0001-fix-ver-gen-thresholds.patch
+
+Fixes VER_GEN12 (1,000,000 → 1,200) and VER_GEN13 (1,001,030 → 1,300).
+
+The original VER_GEN12 value was an unreachable placeholder that caused all Intel
+Arc GPUs (cc≈1255 for A770) to fall through to the NVIDIA Ampere tuning path in
+all MMQ kernels. After this patch, Intel discrete GPUs use the VER_GEN12 path.
+
+## 0002-tune-dmmv-xy-for-arc.patch
+
+Changes presets.hpp: DMMV_X 32→64, MMV_Y 1→2.
+
+Doubles the data processed per thread in DMMV kernels and doubles rows per
+work-group. All common model widths (4096-14336) are divisible by 64.
+
+## 0003-tune-dmmv-xy-common-hpp.patch
+
+Same changes as 0002 but in common.hpp (duplicate definitions).
+
+### Expected impact
+5-15% additional improvement on top of Phase 1.
+
+### ⚠️ Needs Benchmarking
+DMMV_X=64 and MMV_Y=2 were chosen analytically, not empirically. If MMV_Y=2
+causes register spills (check with `GGML_SYCL_DEBUG=1`), revert 0002+0003 and
+try DMMV_X=64 with MMV_Y=1 only.
+
+### Testing checklist
+- [ ] Build succeeds
+- [ ] Unit tests pass
+- [ ] Dense model inference produces correct output
+- [ ] No assertion failures (`ncols % GGML_SYCL_DMMV_X == 0`)
+- [ ] Benchmark comparison vs Phase 1 only

repos/patch/phase3-vulkan-intel/0001-arc-140t-xe2-override.patch

@@ -0,0 +1,34 @@
+diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+index b2a54bd..3469f6f 100644
+--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
++++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+@@ -302,6 +302,14 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
+             return vk_device_architecture::OTHER;
+         }
+ 
++        // Arrow Lake H (Arc 140T) misreports minSubgroupSize=8 despite being Xe2.
++        // Device-ID override to force Xe2 classification.
++        // See: https://github.com/ggml-org/llama.cpp/issues/20776
++        const uint32_t devid = props.deviceID;
++        if (devid == 0x7D51 || devid == 0x7D45) {
++            return vk_device_architecture::INTEL_XE2;
++        }
++
+         vk::PhysicalDeviceProperties2 props2;
+         vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd;
+         vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
+@@ -341,6 +349,14 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
+             return vk_device_architecture::OTHER;
+         }
+ 
++        // Arrow Lake H (Arc 140T) misreports minSubgroupSize=8 despite being Xe2.
++        // Device-ID override to force Xe2 classification.
++        // See: https://github.com/ggml-org/llama.cpp/issues/20776
++        const uint32_t devid = props.deviceID;
++        if (devid == 0x7D51 || devid == 0x7D45) {
++            return vk_device_architecture::INTEL_XE2;
++        }
++
+         vk::PhysicalDeviceProperties2 props2;
+         vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+ 

repos/patch/phase3-vulkan-intel/README.md

@@ -0,0 +1,30 @@
+# Phase 3 — Vulkan Intel Fixes
+
+**Depends on:** Phase 1 and 2 (should be applied and tested first)
+
+## 0001-arc-140t-xe2-override.patch
+
+Adds device-ID override for Intel Arc 140T (Arrow Lake H) to force INTEL_XE2
+classification in the Vulkan backend.
+
+### Problem
+Arc 140T reports minSubgroupSize=8 instead of 16. The Vulkan backend uses
+minSubgroupSize to detect Xe2. When misreported, the 140T is classified as
+OTHER, disabling cooperative matrix and all dependent optimizations.
+
+### Fix
+Checks for Arrow Lake H device IDs (0x7D51, 0x7D45) before the minSubgroupSize
+check and returns INTEL_XE2 directly.
+
+### Applies to
+Both the EXT and KHR code paths in ggml-vulkan.cpp.
+
+### Impact
+Only affects Arrow Lake H (Arc 140T) systems. No effect on other hardware.
+
+### Testing checklist
+- [ ] Build succeeds with Vulkan support
+- [ ] Arc 140T: device classified as INTEL_XE2
+- [ ] Arc 140T: cooperative matrix shaders used for matmul
+- [ ] Other Intel GPUs: no change in behavior
+- [ ] Non-Intel GPUs: no change in behavior