gpufl 0.1.4__tar.gz → 1.0.0__tar.gz

{gpufl-0.1.4 → gpufl-1.0.0}/.github/workflows/release.yml

@@ -136,16 +136,65 @@ jobs:
         uses: pypa/cibuildwheel@v2.22.0
         env:
           CIBW_VIRTUALENV_VERSION: "20.27.1"
-          CIBW_ENVIRONMENT_LINUX: "CUDA_HOME=/usr/local/cuda PATH=/usr/local/cuda/bin:$PATH CMAKE_ARGS='-DGPUFL_ENABLE_NVIDIA=ON -DGPUFL_ENABLE_AMD=OFF -DBUILD_TESTING=OFF'"
+          # OpenSSL hints: manylinux_2_28 (AlmaLinux 8) ships OpenSSL 1.1.1
+          # as the system default, but cpp-httplib v0.18.5 requires >= 3.0.0
+          # (SSL_get1_peer_certificate). We install EL8's `openssl3-devel`
+          # compat package (see CIBW_BEFORE_ALL_LINUX) which lays OpenSSL 3.x
+          # down under NON-standard prefixes — headers in /usr/include/openssl3,
+          # dev symlinks in /usr/lib64/openssl3 — so find_package(OpenSSL)
+          # won't see it without these explicit cache vars. The runtime
+          # SONAME is still libssl.so.3 in /usr/lib64, so auditwheel bundles
+          # it into gpufl.libs/ as before. Verified against the actual
+          # quay.io/pypa/manylinux_2_28_x86_64 image (OpenSSL 3.5.5).
+          CIBW_ENVIRONMENT_LINUX: "CUDA_HOME=/usr/local/cuda PATH=/usr/local/cuda/bin:$PATH CMAKE_ARGS='-DGPUFL_ENABLE_NVIDIA=ON -DGPUFL_ENABLE_AMD=OFF -DBUILD_TESTING=OFF -DOPENSSL_INCLUDE_DIR=/usr/include/openssl3 -DOPENSSL_SSL_LIBRARY=/usr/lib64/openssl3/libssl.so -DOPENSSL_CRYPTO_LIBRARY=/usr/lib64/openssl3/libcrypto.so'"
+          # Windows build needs the OpenSSL install path so find_package(OpenSSL)
+          # in CMakeLists.txt succeeds, otherwise HTTPS upload (HttpLogSink)
+          # silently falls back to HTTP-only — see openssl-windows.html in the
+          # manual repo for the user-facing story. CIBW_BEFORE_ALL_WINDOWS
+          # installs choco's openssl package into this path.
+          CIBW_ENVIRONMENT_WINDOWS: >-
+            OPENSSL_ROOT_DIR="C:/Program Files/OpenSSL-Win64"
+            CMAKE_ARGS="-DGPUFL_ENABLE_NVIDIA=ON -DGPUFL_ENABLE_AMD=OFF -DBUILD_TESTING=OFF"
           # cuda-nvml-devel-13-1 ships the libnvidia-ml.so stub under
           # targets/x86_64-linux/lib/stubs/ — without it CMake's NVML probe
           # finds nothing and (since v0.1.1) fails the build loudly. Every
           # release before v0.1.1 silently shipped wheels without NVML
           # because this package was missing here.
+          #
+          # openssl3-devel (NOT openssl-devel — that's 1.1.1 on EL8) provides
+          # OpenSSL 3.x headers + .so symlinks so cpp-httplib's
+          # CPPHTTPLIB_OPENSSL_SUPPORT path compiles (it #errors on < 3.0.0).
+          # It installs under /usr/include/openssl3 + /usr/lib64/openssl3 —
+          # see the OPENSSL_* hints in CIBW_ENVIRONMENT_LINUX. auditwheel
+          # bundles the resulting libssl.so.3 / libcrypto.so.3 (from
+          # /usr/lib64) into the wheel under gpufl.libs/ automatically
+          # (they're not on the manylinux_2_28 whitelist or our --exclude list).
           CIBW_BEFORE_ALL_LINUX: >-
             curl -fsSL https://developer.download.nvidia.com/compute/cuda/repos/rhel8/x86_64/cuda-rhel8.repo > /etc/yum.repos.d/cuda.repo &&
-            dnf install -y --nogpgcheck cuda-nvcc-13-1 cuda-cudart-devel-13-1 cuda-cupti-13-1 cuda-driver-devel-13-1 cuda-nvml-devel-13-1
-          CIBW_MANYLINUX_X86_64_IMAGE: manylinux_2_28
+            dnf install -y --nogpgcheck cuda-nvcc-13-1 cuda-cudart-devel-13-1 cuda-cupti-13-1 cuda-driver-devel-13-1 cuda-nvml-devel-13-1 openssl3-devel
+          # Install OpenSSL on the Windows runner so find_package(OpenSSL)
+          # in CMakeLists.txt succeeds and cpp-httplib gets compiled with
+          # CPPHTTPLIB_OPENSSL_SUPPORT=1. Chocolatey is pre-installed on
+          # GitHub's windows-latest runners; the openssl package installs
+          # to C:\Program Files\OpenSSL-Win64\ (matches OPENSSL_ROOT_DIR
+          # in CIBW_ENVIRONMENT_WINDOWS above).
+          CIBW_BEFORE_ALL_WINDOWS: choco install -y openssl --no-progress
+          # cibuildwheel ships NO default Windows repair command and only
+          # auto-installs delvewheel for that (nonexistent) default. Because
+          # we override CIBW_REPAIR_WHEEL_COMMAND_WINDOWS below (to pass
+          # --add-path for the OpenSSL DLLs), we must install delvewheel
+          # ourselves. The repair step runs in this same build env, so the
+          # tool lands on PATH. Without this the repair dies with
+          # "'delvewheel' is not recognized as an internal or external command".
+          CIBW_BEFORE_BUILD_WINDOWS: pip install delvewheel
+          # Pin a recent manylinux_2_28 image (AlmaLinux 8.10). cibuildwheel
+          # 2.22.0's default pin (2024.11.16-1) is an AlmaLinux 8.6 snapshot
+          # whose repos only carry OpenSSL 1.1.1 — there is NO openssl3
+          # package — so the build died with "No match for argument:
+          # openssl3-devel". 8.10 ships openssl3 (3.5.5). Same glibc-2.28 /
+          # manylinux_2_28 ABI; verified to have cp312+cp313, openssl3-devel,
+          # and libssl.so.3 in /usr/lib64 (so auditwheel still bundles it).
+          CIBW_MANYLINUX_X86_64_IMAGE: quay.io/pypa/manylinux_2_28_x86_64:2026.05.17-1
           CIBW_BUILD: "cp312-manylinux_x86_64 cp313-manylinux_x86_64 cp312-win_amd64 cp313-win_amd64"
           # libnvidia-ml.so.1 is excluded for the same reason as libcuda.so.1:
           # it ships with the NVIDIA driver, not the CUDA toolkit, and is
@@ -156,7 +205,39 @@ jobs:
           # could not be located"). The toolkit's `libnvidia-ml.so` stub is
           # only the unversioned link-time placeholder — the versioned
           # `.so.1` the SONAME chains to lives on the user's machine.
+          #
+          # libssl/libcrypto are NOT excluded — auditwheel bundles them
+          # under gpufl.libs/ so the wheel ships its own OpenSSL and
+          # HTTPS works on user machines without any system install.
           CIBW_REPAIR_WHEEL_COMMAND_LINUX: "auditwheel repair --plat manylinux_2_28_x86_64 --exclude libcuda.so.1 --exclude libnvidia-ml.so.1 -w {dest_dir} {wheel}"
+          # On Windows, cibuildwheel's default is `delvewheel repair`.
+          # We need delvewheel to find the OpenSSL DLLs (libssl-3-x64.dll,
+          # libcrypto-3-x64.dll) so it copies them into the wheel. The
+          # choco install puts them under C:\Program Files\OpenSSL-Win64\bin\
+          # — give that to delvewheel via --add-path. Without this, the
+          # rebuilt wheel imports cleanly on a system that already has
+          # OpenSSL on PATH but fails on a clean machine.
+          # delvewheel vendors the wheel's DLL deps. We must:
+          #  * --add-path the dirs holding the DLLs to bundle. OpenSSL is in
+          #    its choco bin; cudart64_*.dll is in CUDA\vX.Y\bin (on PATH, but
+          #    listed for safety); cupti64_*.dll lives in CUDA's
+          #    extras\CUPTI\lib64, which is NOT on PATH — without it delvewheel
+          #    fails with "Unable to find library: cupti64_2025.4.0.dll". We
+          #    bundle cudart+cupti so the wheel is self-contained, matching the
+          #    Linux wheel (auditwheel bundles libcudart/libcupti). The CUPTI
+          #    version-suffixed DLL name also makes excluding it fragile across
+          #    CUDA point releases, so bundling is the robust choice.
+          #  * --exclude the driver DLLs that ship with the user's driver, not
+          #    the toolkit, and are absent on the GPU-less runner: nvcuda.dll
+          #    (== libcuda.so.1) and nvml.dll (== libnvidia-ml.so.1). Mirrors
+          #    the Linux auditwheel --exclude flags.
+          # Paths use the pinned CUDA 13.1 location (see the Jimver cuda-toolkit
+          # step). Both --add-path and --exclude are ';'-delimited.
+          CIBW_REPAIR_WHEEL_COMMAND_WINDOWS: >-
+            delvewheel repair
+            --add-path "C:\\Program Files\\OpenSSL-Win64\\bin;C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v13.1\\bin;C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v13.1\\extras\\CUPTI\\lib64"
+            --exclude "nvcuda.dll;nvml.dll"
+            -w {dest_dir} {wheel}
 
       - uses: actions/upload-artifact@v4
         with:

{gpufl-0.1.4 → gpufl-1.0.0}/CMakeLists.txt

@@ -1,7 +1,7 @@
 cmake_minimum_required(VERSION 3.31)
 
 project(gpufl_client
-    VERSION 0.1.4
+    VERSION 1.0.0
     LANGUAGES CXX
     DESCRIPTION "Header-only GPU monitoring client library"
 )

{gpufl-0.1.4 → gpufl-1.0.0}/Dockerfile.monitor

@@ -5,10 +5,14 @@
 #
 # Run:
 #   docker run --gpus all \
-#     -e GPUFL_HTTP_URL=http://my-backend:8080/api/v1/events/ \
+#     -e GPUFL_HTTP_HOST=https://api.gpuflight.com \
 #     -e GPUFL_HTTP_TOKEN=gfl_... \
 #     gpufl/monitor:latest
 #
+# GPUFL_HTTP_HOST is just the scheme+host. The agent appends the
+# /api/{version}/events/<type> path automatically; override the
+# version with GPUFL_HTTP_API_VERSION when the backend cuts v2 etc.
+#
 # The Java agent JAR is pulled from the pre-built ghcr.io/gpu-flight/gpufl-agent image.
 # No local gpufl-agent checkout is required.
 

{gpufl-0.1.4 → gpufl-1.0.0}/Dockerfile.monitor.amd

@@ -7,10 +7,14 @@
 #   docker run -d \
 #     --device /dev/kfd --device /dev/dri \
 #     --group-add video --group-add render \
-#     -e GPUFL_HTTP_URL=http://my-backend:8080/api/v1/events/ \
+#     -e GPUFL_HTTP_HOST=https://api.gpuflight.com \
 #     -e GPUFL_HTTP_TOKEN=gfl_... \
 #     gpufl/monitor-amd:latest
 #
+# GPUFL_HTTP_HOST is just the scheme+host. The agent appends the
+# /api/{version}/events/<type> path automatically; override the
+# version with GPUFL_HTTP_API_VERSION when the backend cuts v2 etc.
+#
 # The Java agent JAR is pulled from the pre-built ghcr.io/gpu-flight/gpufl-agent image.
 # No local gpufl-agent checkout is required.
 

{gpufl-0.1.4 → gpufl-1.0.0}/Dockerfile.monitor.supervisord.conf

@@ -21,7 +21,9 @@ stderr_logfile_maxbytes=0
 ;   GPUFL_SOURCE_FOLDER  — must match the log dir used by gpufl-monitor
 ;   GPUFL_SOURCE_PREFIX  — must match GPUFL_MONITOR_LOG_DIR base name (default: session)
 ;   GPUFL_PUBLISHER_TYPE — http or kafka
-;   GPUFL_HTTP_URL       — e.g. http://backend:8080/api/v1/events/
+;   GPUFL_HTTP_HOST      — scheme+host, e.g. https://api.gpuflight.com
+;                          (agent appends /api/{version}/events/<type> automatically)
+;   GPUFL_HTTP_API_VERSION — optional; defaults to v1
 ;   GPUFL_HTTP_TOKEN     — Bearer token
 ;   GPUFL_LOG_TYPES      — default: device,scope,system (override to restrict channels)
 ;   GPUFL_CURSOR_FILE    — default: ./cursor.json (override for persistence across restarts)

{gpufl-0.1.4 → gpufl-1.0.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: gpufl
-Version: 0.1.4
+Version: 1.0.0
 Summary: GPU Monitoring Client
 Author-Email: Myoungho Shin <myounghoshin84@gmail.com>
 Classifier: Development Status :: 3 - Alpha
@@ -22,7 +22,7 @@ Requires-Dist: jax>=0.4; extra == "jax"
 Provides-Extra: triton
 Requires-Dist: triton>=2.1; extra == "triton"
 Provides-Extra: numba
-Requires-Dist: numba; extra == "numba"
+Requires-Dist: numba-cuda; extra == "numba"
 Requires-Dist: numpy; extra == "numba"
 Provides-Extra: viz
 Requires-Dist: pandas>=1.5; extra == "viz"
@@ -36,7 +36,7 @@ Requires-Dist: requests>=2.28; extra == "all"
 Requires-Dist: cupy-cuda12x>=12; extra == "all"
 Requires-Dist: jax>=0.4; extra == "all"
 Requires-Dist: triton>=2.1; extra == "all"
-Requires-Dist: numba; extra == "all"
+Requires-Dist: numba-cuda; extra == "all"
 Requires-Dist: numpy; extra == "all"
 Requires-Dist: pandas>=1.5; extra == "all"
 Requires-Dist: matplotlib>=3.7; extra == "all"
@@ -63,7 +63,7 @@ To keep the initial design coherent, **we are not currently accepting major feat
 
 Try the portal with real session data — no sign-up required:
 
-**[https://gpufl-front.vercel.app/demo/gdemo_4R98GA5MzYdosqvNsUqdp_MaUgEcDHABS2C5PHbCDQE](https://gpufl-front.vercel.app/demo/gdemo_4R98GA5MzYdosqvNsUqdp_MaUgEcDHABS2C5PHbCDQE)**
+**[Demo Link](https://demo.gpuflight.com)**
 
 ## Key Features
 
@@ -128,7 +128,6 @@ import gpufl
 gpufl.init("my-app",
            log_path="./my_logs",
            sampling_auto_start=True,
-           enable_kernel_details=True,
            enable_stack_trace=True)
 
 a = torch.randn(1024, 1024, device="cuda")
@@ -150,8 +149,7 @@ from gpufl import ProfilingEngine
 
 gpufl.init("my-app",
            log_path="./logs",
-           profiling_engine=ProfilingEngine.PcSampling,
-           enable_kernel_details=True)
+           profiling_engine=ProfilingEngine.PcSampling)
 ```
 
 | Engine | What it collects | Analyzer method | Best for |
@@ -167,7 +165,6 @@ gpufl.init("my-app",
 gpufl::InitOptions opts;
 opts.app_name = "my_app";
 opts.log_path = "my_logs";
-opts.enable_kernel_details = true;
 opts.enable_stack_trace = true;
 opts.sampling_auto_start = true;
 opts.profiling_engine = gpufl::ProfilingEngine::SassMetrics;
@@ -305,6 +302,60 @@ viz.show()
 
 ---
 
+## Report Generation
+
+For a quick, shareable text summary of a session — session metadata, kernel
+hotspots, duration percentiles, and system metrics — generate a **text report**.
+It's the fastest way to see "what happened" without opening the dashboard, and
+it drops cleanly into CI logs, PR comments, or a plain terminal.
+
+![Text report example](images/Screenshot2.png)
+
+The report includes:
+- **Session Summary** — app name, session ID, duration, GPU device + SM count.
+- **Kernel Execution Summary** — total / unique kernels, GPU-busy %, and
+  duration statistics (avg / median / P90 / P99 / min / max). When a SASS
+  profiling engine was active, kernel durations include instrumentation
+  overhead and the report labels them accordingly.
+- **Top kernels by total GPU time** — with per-kernel call counts.
+- **Per-kernel details** — grid/block dimensions, occupancy, registers,
+  shared memory (static + dynamic), register spills, and Waves/SM.
+
+### From C++
+
+Call `generateReport()` after `shutdown()` — it reads the NDJSON logs written
+during the session:
+
+```cpp
+gpufl::init(opts);
+// ... your CUDA / HIP work ...
+gpufl::shutdown();
+
+gpufl::generateReport();               // print to stdout
+gpufl::generateReport("report.txt");   // or save to a file
+```
+
+### From Python
+
+```python
+from gpufl.report import generate_report
+
+# Print the report — wrap in print() so newlines render. In a Jupyter
+# notebook this also keeps the table columns aligned (stdout renders in
+# a monospace font). A bare `generate_report(...)` as a cell's last
+# expression shows an escaped one-line string, so always print() it.
+text = generate_report("./logs", log_prefix="my_app", top_n=10)
+print(text)
+
+# Or save it straight to a file
+generate_report("./logs", log_prefix="my_app", top_n=10, output_path="report.txt")
+```
+
+The Python version reads the same NDJSON logs the analyzer uses — no GPU
+required, so you can generate reports from logs copied off another machine.
+
+---
+
 ## Testing
 
 ### C++ Tests
@@ -345,5 +396,13 @@ To allow non-root users to profile GPU kernels (using CUPTI/PC Sampling) on Linu
 
 ---
 
-*GPU Flight is open source: [github.com/gpu-flight](https://github.com/gpu-flight)*
-*Python package: [pypi.org/project/gpufl](https://pypi.org/project/gpufl/)*
+## Where your logs go
+
+By default the client writes NDJSON to disk. To stream them to a hosted
+dashboard, set `backend_url` + `api_key` (or the `GPUFL_BACKEND_URL` /
+`GPUFL_API_KEY` env vars) and they're delivered live to
+[app.gpuflight.com](https://app.gpuflight.com). Create a workspace at
+[gpuflight.com](https://gpuflight.com)
+
+This client (gpufl-client) is open source. The ingestion service and
+the dashboard UI are proprietary and managed-only today.

{gpufl-0.1.4 → gpufl-1.0.0}/README.md

@@ -18,7 +18,7 @@ To keep the initial design coherent, **we are not currently accepting major feat
 
 Try the portal with real session data — no sign-up required:
 
-**[https://gpufl-front.vercel.app/demo/gdemo_4R98GA5MzYdosqvNsUqdp_MaUgEcDHABS2C5PHbCDQE](https://gpufl-front.vercel.app/demo/gdemo_4R98GA5MzYdosqvNsUqdp_MaUgEcDHABS2C5PHbCDQE)**
+**[Demo Link](https://demo.gpuflight.com)**
 
 ## Key Features
 
@@ -83,7 +83,6 @@ import gpufl
 gpufl.init("my-app",
            log_path="./my_logs",
            sampling_auto_start=True,
-           enable_kernel_details=True,
            enable_stack_trace=True)
 
 a = torch.randn(1024, 1024, device="cuda")
@@ -105,8 +104,7 @@ from gpufl import ProfilingEngine
 
 gpufl.init("my-app",
            log_path="./logs",
-           profiling_engine=ProfilingEngine.PcSampling,
-           enable_kernel_details=True)
+           profiling_engine=ProfilingEngine.PcSampling)
 ```
 
 | Engine | What it collects | Analyzer method | Best for |
@@ -122,7 +120,6 @@ gpufl.init("my-app",
 gpufl::InitOptions opts;
 opts.app_name = "my_app";
 opts.log_path = "my_logs";
-opts.enable_kernel_details = true;
 opts.enable_stack_trace = true;
 opts.sampling_auto_start = true;
 opts.profiling_engine = gpufl::ProfilingEngine::SassMetrics;
@@ -260,6 +257,60 @@ viz.show()
 
 ---
 
+## Report Generation
+
+For a quick, shareable text summary of a session — session metadata, kernel
+hotspots, duration percentiles, and system metrics — generate a **text report**.
+It's the fastest way to see "what happened" without opening the dashboard, and
+it drops cleanly into CI logs, PR comments, or a plain terminal.
+
+![Text report example](images/Screenshot2.png)
+
+The report includes:
+- **Session Summary** — app name, session ID, duration, GPU device + SM count.
+- **Kernel Execution Summary** — total / unique kernels, GPU-busy %, and
+  duration statistics (avg / median / P90 / P99 / min / max). When a SASS
+  profiling engine was active, kernel durations include instrumentation
+  overhead and the report labels them accordingly.
+- **Top kernels by total GPU time** — with per-kernel call counts.
+- **Per-kernel details** — grid/block dimensions, occupancy, registers,
+  shared memory (static + dynamic), register spills, and Waves/SM.
+
+### From C++
+
+Call `generateReport()` after `shutdown()` — it reads the NDJSON logs written
+during the session:
+
+```cpp
+gpufl::init(opts);
+// ... your CUDA / HIP work ...
+gpufl::shutdown();
+
+gpufl::generateReport();               // print to stdout
+gpufl::generateReport("report.txt");   // or save to a file
+```
+
+### From Python
+
+```python
+from gpufl.report import generate_report
+
+# Print the report — wrap in print() so newlines render. In a Jupyter
+# notebook this also keeps the table columns aligned (stdout renders in
+# a monospace font). A bare `generate_report(...)` as a cell's last
+# expression shows an escaped one-line string, so always print() it.
+text = generate_report("./logs", log_prefix="my_app", top_n=10)
+print(text)
+
+# Or save it straight to a file
+generate_report("./logs", log_prefix="my_app", top_n=10, output_path="report.txt")
+```
+
+The Python version reads the same NDJSON logs the analyzer uses — no GPU
+required, so you can generate reports from logs copied off another machine.
+
+---
+
 ## Testing
 
 ### C++ Tests
@@ -300,5 +351,13 @@ To allow non-root users to profile GPU kernels (using CUPTI/PC Sampling) on Linu
 
 ---
 
-*GPU Flight is open source: [github.com/gpu-flight](https://github.com/gpu-flight)*
-*Python package: [pypi.org/project/gpufl](https://pypi.org/project/gpufl/)*
+## Where your logs go
+
+By default the client writes NDJSON to disk. To stream them to a hosted
+dashboard, set `backend_url` + `api_key` (or the `GPUFL_BACKEND_URL` /
+`GPUFL_API_KEY` env vars) and they're delivered live to
+[app.gpuflight.com](https://app.gpuflight.com). Create a workspace at
+[gpuflight.com](https://gpuflight.com)
+
+This client (gpufl-client) is open source. The ingestion service and
+the dashboard UI are proprietary and managed-only today.

{gpufl-0.1.4 → gpufl-1.0.0}/daemon/README.md

@@ -79,7 +79,7 @@ docker build \
 Copy `.env.example` to `.env` and set the required variables, then:
 
 ```bash
-GPUFL_HTTP_URL=https://your-backend/api/v1/events/ \
+GPUFL_HTTP_HOST=https://your-backend \
 GPUFL_HTTP_TOKEN=gfl_your_token_here \
 docker compose -f docker-compose.monitor.yml up -d
 ```
@@ -99,7 +99,7 @@ docker compose -f docker-compose.monitor.yml down
 ### AMD
 
 ```bash
-GPUFL_HTTP_URL=https://your-backend/api/v1/events/ \
+GPUFL_HTTP_HOST=https://your-backend \
 GPUFL_HTTP_TOKEN=gfl_your_token_here \
 docker compose -f docker-compose.monitor.amd.yml up -d
 ```
@@ -127,7 +127,7 @@ docker run -d \
   --name gpufl-monitor \
   --gpus all \
   --restart unless-stopped \
-  -e GPUFL_HTTP_URL=https://your-backend/api/v1/events/ \
+  -e GPUFL_HTTP_HOST=https://your-backend \
   -e GPUFL_HTTP_TOKEN=gfl_your_token_here \
   -v gpufl-cursor:/var/gpufl/monitor \
   gpufl/monitor:latest
@@ -141,7 +141,7 @@ docker run -d \
   --device /dev/kfd --device /dev/dri \
   --group-add video --group-add render \
   --restart unless-stopped \
-  -e GPUFL_HTTP_URL=https://your-backend/api/v1/events/ \
+  -e GPUFL_HTTP_HOST=https://your-backend \
   -e GPUFL_HTTP_TOKEN=gfl_your_token_here \
   -v gpufl-cursor-amd:/var/gpufl/monitor \
   gpufl/monitor-amd:latest
@@ -175,7 +175,8 @@ The named volume persists the agent's read cursor so it resumes from where it le
 | Variable | Default | Description |
 |---|---|---|
 | `GPUFL_PUBLISHER_TYPE` | `http` | Publisher backend: `http` or `kafka` |
-| `GPUFL_HTTP_URL` | *(required)* | Backend ingest URL, e.g. `https://app.gpuflight.io/api/v1/events/` |
+| `GPUFL_HTTP_HOST` | *(required)* | Backend scheme+host, e.g. `https://api.gpuflight.com`. The agent appends `/api/{version}/events/<type>` automatically. |
+| `GPUFL_HTTP_API_VERSION` | `v1` | Backend API version. Bump when the backend cuts v2 etc. |
 | `GPUFL_HTTP_TOKEN` | *(empty)* | Bearer token for the backend API |
 | `GPUFL_HTTP_TIMEOUT_SEC` | `10` | HTTP request timeout in seconds |
 

{gpufl-0.1.4 → gpufl-1.0.0}/docker-compose.monitor.amd.yml

@@ -22,9 +22,13 @@ services:
       GPUFL_LOG_TYPES: ${GPUFL_LOG_TYPES:-device,scope,system}
       GPUFL_CURSOR_FILE: ${GPUFL_CURSOR_FILE:-/var/gpufl/monitor/cursor.json}
 
-      # Java agent — publisher (HTTP)
+      # Java agent — publisher (HTTP). Set GPUFL_HTTP_HOST to just the
+      # scheme+host (e.g. https://api.gpuflight.com); the agent builds
+      # the /api/{version}/events/<type> path itself. Override the
+      # version via GPUFL_HTTP_API_VERSION when the backend bumps to v2.
       GPUFL_PUBLISHER_TYPE: ${GPUFL_PUBLISHER_TYPE:-http}
-      GPUFL_HTTP_URL: ${GPUFL_HTTP_URL}
+      GPUFL_HTTP_HOST: ${GPUFL_HTTP_HOST}
+      GPUFL_HTTP_API_VERSION: ${GPUFL_HTTP_API_VERSION:-v1}
       GPUFL_HTTP_TOKEN: ${GPUFL_HTTP_TOKEN:-}
       GPUFL_HTTP_TIMEOUT_SEC: ${GPUFL_HTTP_TIMEOUT_SEC:-10}
 

{gpufl-0.1.4 → gpufl-1.0.0}/docker-compose.monitor.yml

@@ -39,9 +39,13 @@ services:
       GPUFL_SOURCE_FOLDERS: ${GPUFL_SOURCE_FOLDERS:-/var/gpufl/monitor,/var/gpufl/demo}
       GPUFL_CURSOR_FILE: ${GPUFL_CURSOR_FILE:-/var/gpufl/monitor/cursor.json}
 
-      # Java agent — publisher (HTTP)
+      # Java agent — publisher (HTTP). Set GPUFL_HTTP_HOST to just the
+      # scheme+host (e.g. https://api.gpuflight.com); the agent builds
+      # the /api/{version}/events/<type> path itself. Override the
+      # version via GPUFL_HTTP_API_VERSION when the backend bumps to v2.
       GPUFL_PUBLISHER_TYPE: ${GPUFL_PUBLISHER_TYPE:-http}
-      GPUFL_HTTP_URL: ${GPUFL_HTTP_URL}
+      GPUFL_HTTP_HOST: ${GPUFL_HTTP_HOST}
+      GPUFL_HTTP_API_VERSION: ${GPUFL_HTTP_API_VERSION:-v1}
       GPUFL_HTTP_TOKEN: ${GPUFL_HTTP_TOKEN:-}
       GPUFL_HTTP_TIMEOUT_SEC: ${GPUFL_HTTP_TIMEOUT_SEC:-10}
 

{gpufl-0.1.4 → gpufl-1.0.0}/example/amd/gpufl_scope_demo.cpp

@@ -130,7 +130,6 @@ int main() {
     opts.system_sample_rate_ms = 50;
     opts.kernel_sample_rate_ms = 0;
     opts.sampling_auto_start = true;
-    opts.enable_kernel_details = true;
     opts.enable_debug_output = true;
     opts.enable_stack_trace = false;
     opts.profiling_engine = gpufl::ProfilingEngine::SassMetrics;

{gpufl-0.1.4 → gpufl-1.0.0}/example/cuda/block_style_example.cu

@@ -38,7 +38,6 @@ int main() {
     opts.log_path = "gfl_block";
     opts.system_sample_rate_ms = 50;
     opts.kernel_sample_rate_ms = 50;
-    opts.enable_kernel_details = true;
     opts.sampling_auto_start = true;
     opts.enable_debug_output = true;
     opts.enable_source_collection = true;

{gpufl-0.1.4 → gpufl-1.0.0}/example/cuda/memory_coalescing_demo.cu

@@ -61,7 +61,6 @@ int main() {
     opts.log_path = "memory_coalescing_demo";
     opts.system_sample_rate_ms = 10;
     opts.kernel_sample_rate_ms = 10;
-    opts.enable_kernel_details = true;
     opts.sampling_auto_start = true;
     opts.enable_debug_output = true;
     opts.profiling_engine = gpufl::ProfilingEngine::PcSamplingWithSass;

{gpufl-0.1.4 → gpufl-1.0.0}/example/cuda/occupancy_demo.cu

@@ -79,7 +79,6 @@ int main()
     gpufl::InitOptions opts;
     opts.app_name              = "occupancy_demo";
     opts.log_path              = "occupancy_demo.log";
-    opts.enable_kernel_details = true;  // required for occupancy breakdown fields
     opts.sampling_auto_start   = true;
     opts.enable_debug_output   = false;
 

{gpufl-0.1.4 → gpufl-1.0.0}/example/cuda/sass_divergence_demo.cu

@@ -144,15 +144,12 @@ int main() {
     gpufl::InitOptions opts;
     opts.app_name = "sass_divergence_demo";
     opts.log_path = "sass_divergence";
-    if (const char* k = std::getenv("GPUFL_API_KEY")) opts.api_key = k;
-    opts.backend_url = "http://localhost:8080";
-    opts.remote_upload = true;
+    opts.remote_upload = false;
     opts.system_sample_rate_ms = 10;
-    opts.enable_kernel_details = true;
     opts.enable_debug_output = true;
     opts.sampling_auto_start = true;
     opts.enable_stack_trace = true;
-    opts.profiling_engine = gpufl::ProfilingEngine::PcSampling;
+    opts.profiling_engine = gpufl::ProfilingEngine::PcSamplingWithSass;
 
     if (!gpufl::init(opts)) {
         std::cerr << "Failed to initialize gpufl" << std::endl;

{gpufl-0.1.4 → gpufl-1.0.0}/example/cuda/vector_add_benchmark.cu

@@ -27,7 +27,6 @@ int main() {
     opts.log_path = "vector_add_benchmark";
     opts.system_sample_rate_ms = 50;
     opts.kernel_sample_rate_ms = 50;
-    opts.enable_kernel_details = true;
     opts.sampling_auto_start = true;
     opts.enable_debug_output = true;
     opts.enable_source_collection = true;

{gpufl-0.1.4 → gpufl-1.0.0}/example/python/02_numba_cuda.py

@@ -1,7 +1,9 @@
 import gpufl as gfl
+from gpufl.report import generate_report
 import numpy as np
 from numba import cuda
 import math
+import os
 import time
 
 # --- 1. Define a Real CUDA Kernel (Matrix Mul) ---
@@ -20,9 +22,32 @@ def matmul_kernel(A, B, C):
 
 def run_benchmark():
     # --- 2. Initialize GPUFL ---
-    # We enable the background sampler (16ms) to catch VRAM/Power usage during the heavy compute
+    # LOG_PATH is the file prefix the FileLogSink writes to — it produces
+    # <LOG_PATH>.device.log / .scope.log / .system.log. We reuse it below
+    # to point generate_report() at the same files.
+    LOG_PATH = "./gfl_logs"
+
+    BACKEND_URL = os.environ.get("GPUFL_BACKEND_URL", "api.gpuflight.com")
+    API_KEY = os.environ.get("GPUFL_API_KEY", "")
+    REMOTE_UPLOAD = bool(API_KEY)
+
     print("[GPUFL] Initializing...")
-    gfl.init("Numba_App", "./gfl_logs", 100)
+    if REMOTE_UPLOAD:
+        print(f"[GPUFL] Live upload ON -> {BACKEND_URL}")
+    else:
+        print("[GPUFL] Live upload OFF (set GPUFL_API_KEY to enable). Local files only.")
+
+    gfl.init(
+        app_name="Numba_App",
+        log_path=LOG_PATH,
+        sampling_auto_start=True,
+        system_sample_rate_ms=100,
+        enable_debug_output=True,
+        profiling_engine=gfl.ProfilingEngine.PcSamplingWithSass,
+        backend_url=BACKEND_URL,
+        api_key=API_KEY,
+        remote_upload=REMOTE_UPLOAD,
+    )
 
     try:
         # --- 3. Setup Data (Heavy Load) ---
@@ -70,6 +95,21 @@ def run_benchmark():
         print("[GPUFL] Shutting down...")
         gfl.shutdown()
 
+        # --- 6. Generate a text report from the logs we just wrote ---
+        # shutdown() above flushes and closes the NDJSON channels, so the
+        # report reflects the full session. generate_report reads the same
+        # logs the analyzer uses — no GPU required for this step. We split
+        # LOG_PATH into (dir, prefix) the way GpuFlightSession expects:
+        #   "./gfl_logs" -> dir=".", prefix="gfl_logs"
+        #                -> reads ./gfl_logs.{device,scope,system}.log
+        # Wrap in print() so the report renders with real newlines (and,
+        # in a Jupyter notebook, in the monospace stdout area so the
+        # kernel tables stay aligned).
+        log_dir = os.path.dirname(LOG_PATH) or "."
+        log_prefix = os.path.basename(LOG_PATH)
+        print("\n[GPUFL] Session report:\n")
+        print(generate_report(log_dir, log_prefix=log_prefix, top_n=10))
+
 if __name__ == "__main__":
     if cuda.is_available():
         run_benchmark()

{gpufl-0.1.4 → gpufl-1.0.0}/example/python/03_pytorch_benchmark.py

@@ -31,9 +31,7 @@ def run_stress_test():
                sampling_auto_start=True,
                system_sample_rate_ms=50,
                kernel_sample_rate_ms=50,
-               enable_kernel_details=True,
-               enable_debug_output=True,
-               enable_profiling=True,
+               enable_debug_output=False,
                enable_stack_trace=True,
                # opt-in to memory tracking. Default-off in v1
                # because TF eager and similar workloads can produce
@@ -49,7 +47,7 @@ def run_stress_test():
                remote_upload=remote_upload,
                api_key=api_key,
                backend_url=backend_url,
-               profiling_engine=gpufl.ProfilingEngine.PcSamplingWithSass)
+               profiling_engine=gpufl.ProfilingEngine.PcSampling)
 
     try:
         # 2. Allocate (Uses approx 3GB VRAM)