Add packed INT4 GPTQ decode and page sizes 64 and 128

.dockerignore

@@ -0,0 +1,17 @@
+.git
+.github
+.specify
+specs
+build
+dist
+*.egg-info
+__pycache__
+.venv
+venv
+.env
+.env.*
+*.log
+.DS_Store
+benchmark/*.log
+tests/__pycache__
+*.pyc

.gitignore

@@ -43,11 +43,56 @@
 # python build files
 *.pyc
 *.pyo
+__pycache__/
+*.egg-info/
+.venv/
+venv/
+dist/
+.env
+.env.*
 
 build
 
 # vscode
 .vscode/
+.idea/
+.DS_Store
+*.tmp
+*.swp
 
 # benchmark run logs (benchmark/run_suite.py writes one per benchmark)
 benchmark/*.log
+
+# AI assistant scratch.
+# Spec Kit plans stay local. .claude/skills/ stays tracked.
+# Ignore tool indexes, transcripts, and generation scratch only.
+.specify/
+specs/
+.cursor/
+.cursor-*/
+*.cursor-*
+.cursorrules
+.cursorindexingignore
+agent-transcripts/
+agent-logs/
+*agent-transcript*
+*agent-state*
+*.agent.jsonl
+*.agent.log
+
+# Claude Code local state (not project skills)
+.claude/settings.local.json
+.claude/*.local.json
+.claude/*.local.md
+.claude/local/
+.claude/projects/
+.claude/sessions/
+.claude/cache/
+.claude/debug/
+.claude/todos/
+.claude/shell-snapshots/
+.claude/ide/
+.claude/statsig/
+.claude/transcripts/
+.claude/history.jsonl
+.claude/**/*.jsonl

Dockerfile.xpu

@@ -0,0 +1,55 @@
+# Torch 2.14.0+xpu, oneAPI 2026.1, and SGLang 0.5.22 are already in this image.
+# The checkpoint stays on the node at /tank/nas/models and is not copied here.
+# pip-installing this repo would replace sgl_kernel and drop FMHA and GDN.
+# Those stay in the base image. This image only adds the GPTQ op.
+FROM lmsysorg/sglang:v0.5.22-xpu
+
+COPY include /opt/sgl-kernel-xpu/include
+COPY src/sycl/gptq_gemm.cpp src/sycl/gptq_entry_probe.cpp src/sycl/gptq_register.cpp src/sycl/gdn_verify_entry.cpp src/sycl/causal_conv1d.cpp src/sycl/causal_conv1d_verify.hpp src/sycl/causal_conv1d_token_parallel.hpp src/sycl/Utils.h src/sycl/SYCLHelpers.h src/sycl/SGLKernelPerf.h /opt/sgl-kernel-xpu/src/
+COPY src/sycl/gdn_attn/gated_delta_rule.hpp /opt/sgl-kernel-xpu/src/gdn_attn/gated_delta_rule.hpp
+COPY src/jit/jit_arch.h /opt/sgl-kernel-xpu/src/jit/jit_arch.h
+
+RUN set -eux; \
+    ICPX=/opt/intel/oneapi/compiler/2026.1/bin/icpx; \
+    TORCH="$(python -c 'import pathlib, torch; print(pathlib.Path(torch.__file__).parent)')"; \
+    "$ICPX" -fsycl -fsycl-targets=spir64_gen \
+      -Xspirv-translator -spirv-ext=+SPV_INTEL_split_barrier,+SPV_INTEL_2d_block_io,+SPV_INTEL_subgroup_matrix_multiply_accumulate \
+      -Xsycl-target-backend "-device bmg" \
+      -std=c++20 -O2 -fPIC -shared \
+      -I/opt/sgl-kernel-xpu/include \
+      -I/opt/sgl-kernel-xpu/src \
+      -I"$TORCH/include" \
+      -I"$TORCH/include/torch/csrc/api/include" \
+      /opt/sgl-kernel-xpu/src/gptq_register.cpp \
+      /opt/sgl-kernel-xpu/src/gptq_gemm.cpp \
+      /opt/sgl-kernel-xpu/src/gptq_entry_probe.cpp \
+      /opt/sgl-kernel-xpu/src/gdn_verify_entry.cpp \
+      /opt/sgl-kernel-xpu/src/causal_conv1d.cpp \
+      -o /opt/sgl-kernel-xpu/sgl_gptq_ops.so \
+      -L"$TORCH/lib" -Wl,-rpath,"$TORCH/lib" \
+      -ltorch -ltorch_xpu -ltorch_cpu -lc10
+
+COPY python/sgl_kernel/gptq_loader.py /opt/sgl-kernel-xpu/python/gptq_loader.py
+COPY scripts /opt/sgl-kernel-xpu/scripts
+COPY deploy/sglang-gptq-xpu.patch deploy/sglang-qwen35-ba.patch deploy/sglang-qwen35-time.patch deploy/sglang-xpu-page.patch deploy/sglang-mamba-page.patch deploy/sglang-xpu-gdn-page.patch deploy/sglang-lmhead.patch deploy/sglang-loop-time.patch deploy/sglang-page1-graph.patch deploy/sglang-stream-token-ids.patch deploy/sglang-gdn-verify.patch /opt/sgl-kernel-xpu/deploy/
+
+RUN set -eux; \
+    kernel="$(python -c 'import pathlib, sgl_kernel; print(pathlib.Path(sgl_kernel.__file__).parent)')"; \
+    cp /opt/sgl-kernel-xpu/python/gptq_loader.py "$kernel/gptq_loader.py"; \
+    printf '\nfrom sgl_kernel.gptq_loader import gptq_entry_probe, gptq_gemm\n' >> "$kernel/__init__.py"; \
+    site="$(python -c 'import pathlib, sglang; print(pathlib.Path(sglang.__file__).resolve().parent.parent)')"; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-gptq-xpu.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-qwen35-ba.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-qwen35-time.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-xpu-page.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-mamba-page.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-xpu-gdn-page.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-gdn-verify.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-lmhead.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-loop-time.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-page1-graph.patch; \
+    patch -p2 -d "$site" < /opt/sgl-kernel-xpu/deploy/sglang-stream-token-ids.patch; \
+    python /opt/sgl-kernel-xpu/scripts/check_sglang_version.py
+
+WORKDIR /opt/sgl-kernel-xpu
+CMD ["python3", "/opt/sgl-kernel-xpu/scripts/run_gptq_mode.py"]

benchmark/bench_gptq_gemm.py

@@ -0,0 +1,42 @@
+"""Time gptq_gemm on XPU. A machine without XPU prints a skip and is not a B70 result."""
+
+import sys
+
+import torch
+
+
+def main() -> None:
+    if not hasattr(torch, "xpu") or not torch.xpu.is_available():
+        print("XPU unavailable; this run is not an Intel Arc Pro B70 result")
+        return
+    from sgl_kernel import gptq_gemm
+
+    device = torch.device("xpu")
+    k, n = 4096, 4096
+    for m in (1, 16):
+        act = torch.randn(m, k, device=device, dtype=torch.bfloat16)
+        qweight = torch.randint(0, 2**31 - 1, (k // 8, n), device=device, dtype=torch.int32)
+        qzeros = torch.randint(0, 2**31 - 1, (k // 128, n // 8), device=device, dtype=torch.int32)
+        scales = torch.rand(k // 128, n, device=device, dtype=torch.bfloat16) * 0.05 + 0.01
+        g_idx = (torch.arange(k, device=device) // 128).to(torch.int32)
+        for _ in range(3):
+            gptq_gemm(act, qweight, qzeros, scales, g_idx, False, 4, 1)
+        torch.xpu.synchronize()
+        start = torch.xpu.Event(enable_timing=True)
+        end = torch.xpu.Event(enable_timing=True)
+        iters = 10
+        start.record()
+        for _ in range(iters):
+            gptq_gemm(act, qweight, qzeros, scales, g_idx, False, 4, 1)
+        end.record()
+        torch.xpu.synchronize()
+        ms = start.elapsed_time(end) / iters
+        print(f"gptq_gemm M={m} K={k} N={n} {ms:.3f} ms")
+
+
+if __name__ == "__main__":
+    try:
+        main()
+    except Exception as exc:
+        print(f"benchmark did not run: {exc}", file=sys.stderr)
+        raise

deploy/k8s/test-job.yaml

@@ -0,0 +1,68 @@
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: sgl-gptq-xpu-test
+spec:
+  backoffLimit: 0
+  template:
+    spec:
+      restartPolicy: Never
+      nodeName: hal
+      containers:
+        - name: gptq
+          image: ${IMAGE}:${TAG}
+          imagePullPolicy: Always
+          env:
+            - name: MODE
+              value: "${MODE}"
+            - name: STAGE
+              value: "${STAGE}"
+            - name: BASELINE_TPS
+              value: "${BASELINE_TPS}"
+            - name: SGL_GPTQ_PROFILE
+              value: "${SGL_GPTQ_PROFILE}"
+            - name: MAX_NEW_TOKENS
+              value: "${MAX_NEW_TOKENS}"
+            - name: SGL_GPTQ_STEP
+              value: "${SGL_GPTQ_STEP}"
+            - name: SGL_GPTQ_PAGE_SIZE
+              value: "${SGL_GPTQ_PAGE_SIZE}"
+            - name: SGL_GPTQ_CONTEXT_LENGTH
+              value: "${SGL_GPTQ_CONTEXT_LENGTH}"
+            - name: SGL_GPTQ_SPEC
+              value: "${SGL_GPTQ_SPEC}"
+            - name: SGL_GPTQ_SPEC_STEPS
+              value: "${SGL_GPTQ_SPEC_STEPS}"
+            - name: SGL_GPTQ_SPEC_DRAFT_TOKENS
+              value: "${SGL_GPTQ_SPEC_DRAFT_TOKENS}"
+            - name: SGL_GPTQ_MAMBA_CACHE
+              value: "${SGL_GPTQ_MAMBA_CACHE}"
+            - name: SGL_GPTQ_PP
+              value: "${SGL_GPTQ_PP}"
+            - name: SGL_GPTQ_TG
+              value: "${SGL_GPTQ_TG}"
+            - name: SGL_GPTQ_BATCH
+              value: "${SGL_GPTQ_BATCH}"
+            - name: SGL_GPTQ_GRAPH_BS
+              value: "${SGL_GPTQ_GRAPH_BS}"
+            - name: SGL_GPTQ_BATCH_WIDTHS
+              value: "${SGL_GPTQ_BATCH_WIDTHS}"
+            - name: SGL_GPTQ_MAX_RUNNING
+              value: "${SGL_GPTQ_MAX_RUNNING}"
+            - name: SGL_GPTQ_MEM_FRACTION
+              value: "${SGL_GPTQ_MEM_FRACTION}"
+            - name: MODEL_DIR
+              value: "/tank/nas/models/Qwen3.8-27B-TURBO-DF735-882-NM-DAU-GPTQ-Int4-g128-tuned"
+          resources:
+            requests:
+              gpu.intel.com/xe: 1
+            limits:
+              gpu.intel.com/xe: 1
+          volumeMounts:
+            - name: models
+              mountPath: /tank/nas/models
+      volumes:
+        - name: models
+          hostPath:
+            path: /tank/nas/models
+            type: Directory