6.2 dockerfile (#176)

* Trying to modernize the dockerfile, pinning rccl; triton; pytorch; hipblaslt to the latest required versions

* Dockerfile fixes. Using the scaling factors in scaled_mm where they are required by torch 2.5 or acceptable by others

* Building torchvision too when building torch

* gradlib as a not-cmake project doesn't inherit `target_compile_definitions(hipblaslt PUBLIC LEGACY_HIPBLAS_DIRECT )`

* Using a specific torch commit with scaled_mm fix until it is in mainline. Fixed scaled_mm in gradlib for no reason at all

* No point in pinning hipblaslt to rocm6.2 release, if we want to build it, we'll want 0.10

* Removed torch requirement

Dockerfile.rocm

@@ -1,13 +1,14 @@
 # default base image
-ARG BASE_IMAGE="rocm/pytorch:rocm6.1.2_ubuntu20.04_py3.9_pytorch_staging"
+ARG BASE_IMAGE="rocm/pytorch:rocm6.2_ubuntu20.04_py3.9_pytorch_release_2.3.0"
 
 ARG COMMON_WORKDIR=/app
 
 # The following ARGs should be "0" or "1". If "1", the respective component will be built and installed on top of the base image
-ARG BUILD_HIPBLASLT="1"
+ARG BUILD_HIPBLASLT="0"
 ARG BUILD_RCCL="1"
 ARG BUILD_FA="1"
 ARG BUILD_TRITON="1"
+ARG BUILD_PYTORCH="1"
 # This ARG should also be "0" or "1". If "1", the vLLM development directory is obtained via git clone.
 # If "0", it is copied in from the local working directory.
 ARG REMOTE_VLLM="0"
@@ -39,11 +40,12 @@ WORKDIR ${COMMON_WORKDIR}
 # -----------------------
 # hipBLASLt build stages
 FROM base AS build_hipblaslt
-ARG HIPBLASLT_BRANCH="6f65c6e"
-RUN git clone https://github.com/ROCm/hipBLASLt \
+ARG HIPBLASLT_BRANCH="e6da924"
+RUN apt-get purge -y hipblaslt \
+    && git clone https://github.com/ROCm/hipBLASLt.git \
     && cd hipBLASLt \
     && git checkout ${HIPBLASLT_BRANCH} \
-    && SCCACHE_IDLE_TIMEOUT=1800 ./install.sh --architecture ${PYTORCH_ROCM_ARCH} \
+    && SCCACHE_IDLE_TIMEOUT=1800 ./install.sh --architecture ${PYTORCH_ROCM_ARCH} --legacy_hipblas_direct \
     && cd build/release \
     && make package
 FROM scratch AS export_hipblaslt_1
@@ -55,7 +57,7 @@ FROM export_hipblaslt_${BUILD_HIPBLASLT} AS export_hipblaslt
 # -----------------------
 # RCCL build stages
 FROM base AS build_rccl
-ARG RCCL_BRANCH="73221b4"
+ARG RCCL_BRANCH="rocm-6.2.0"
 RUN git clone https://github.com/ROCm/rccl \
     && cd rccl \
     && git checkout ${RCCL_BRANCH} \
@@ -69,7 +71,7 @@ FROM export_rccl_${BUILD_RCCL} AS export_rccl
 # -----------------------
 # flash attn build stages
 FROM base AS build_flash_attn
-ARG FA_BRANCH="ae7928c"
+ARG FA_BRANCH="3cea2fb"
 ARG FA_REPO="https://github.com/ROCm/flash-attention.git"
 RUN git clone ${FA_REPO} \
     && cd flash-attention \
@@ -85,9 +87,9 @@ FROM export_flash_attn_${BUILD_FA} AS export_flash_attn
 # -----------------------
 # Triton build stages
 FROM base AS build_triton
-ARG TRITON_BRANCH="6ddb79b"
-ARG TRITON_REPO="https://github.com/OpenAI/triton.git"
-RUN git clone ${TRITON_REPO} \
+ARG TRITON_BRANCH="e192dba"
+ARG TRITON_REPO="https://github.com/triton-lang/triton.git"
+RUN python3 -m pip install ninja cmake wheel pybind11 && git clone ${TRITON_REPO} \
     && cd triton \
     && git checkout ${TRITON_BRANCH} \
     && cd python \
@@ -105,6 +107,36 @@ RUN cd /opt/rocm/share/amd_smi \
 FROM scratch AS export_amdsmi
 COPY --from=build_amdsmi /opt/rocm/share/amd_smi/dist/*.whl /
 
+FROM base as build_pytorch
+# A commit to fix the output scaling factor issue in _scaled_mm
+# Not yet in 2.5.0-rc1
+ARG PYTORCH_BRANCH="cedc116"
+ARG PYTORCH_VISION_BRANCH="v0.19.1"
+ARG PYTORCH_REPO="https://github.com/ROCm/pytorch.git"
+ARG PYTORCH_VISION_REPO="https://github.com/pytorch/vision.git"
+#RUN --mount=type=bind,from=export_hipblaslt,src=/,target=/install \
+#if ls /install/*.deb; then \
+#    apt-get purge -y hipblaslt \
+#    && dpkg -i /install/*.deb \
+#    && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \
+#    && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status; \
+#fi
+RUN git clone ${PYTORCH_REPO} pytorch \
+    && cd pytorch && git checkout ${PYTORCH_BRANCH} && git submodule update --init --recursive \
+    && python tools/amd_build/build_amd.py \
+    && CMAKE_PREFIX_PATH=$(python3 -c 'import sys; print(sys.prefix)') python3 setup.py bdist_wheel --dist-dir=dist \
+    && pip install dist/*.whl \
+    && cd .. \
+    && git clone ${PYTORCH_VISION_REPO} vision \
+    && cd vision && git checkout ${PYTORCH_VISION_BRANCH} \
+    && python3 setup.py bdist_wheel --dist-dir=dist
+FROM scratch as export_pytorch_1
+ARG COMMON_WORKDIR
+COPY --from=build_pytorch ${COMMON_WORKDIR}/pytorch/dist/*.whl /
+COPY --from=build_pytorch ${COMMON_WORKDIR}/vision/dist/*.whl /
+FROM scratch as export_pytorch_0
+from export_pytorch_${BUILD_PYTORCH} as export_pytorch
+
 # -----------------------
 # vLLM (and gradlib) fetch stages
 FROM base AS fetch_vllm_0
@@ -129,6 +161,11 @@ if ls /install/*.deb; then \
     && sed -i 's/, hipblaslt-dev \(.*\), hipcub-dev/, hipcub-dev/g' /var/lib/dpkg/status \
     && sed -i 's/, hipblaslt \(.*\), hipfft/, hipfft/g' /var/lib/dpkg/status; \
 fi
+# Install pytorch
+RUN --mount=type=bind,from=export_pytorch,src=/,target=/install \
+if ls /install/*.whl; then \
+    pip install /install/*.whl; \
+fi
 # Build vLLM
 RUN cd vllm \
     && python3 setup.py clean --all && python3 setup.py bdist_wheel --dist-dir=dist
@@ -197,6 +234,13 @@ RUN --mount=type=bind,from=export_amdsmi,src=/,target=/install \
     pip uninstall -y amdsmi \
     && pip install /install/*.whl;
 
+RUN --mount=type=bind,from=export_pytorch,src=/,target=/install \
+    if ls /install/*.whl; then \
+        # Preemptively uninstall to prevent pip same-version no-installs
+        pip uninstall -y torch torchvision \
+        && pip install /install/*.whl; \
+    fi
+
 RUN python3 -m pip install --upgrade numba scipy huggingface-hub[cli]
 
 # Install vLLM (and gradlib)

gradlib/gradlib/GemmTuner.py

@@ -15,6 +15,7 @@
 atol = 1
 
 CACHE_INVALIDATE_BUFFERS = int(os.getenv("CACHE_INVALIDATE_BUFFERS", "37"))
+ONE = torch.ones(1, dtype=torch.float32, device='cuda')
 
 
 class Gemm:
@@ -68,6 +69,8 @@ def check_gemm_ref(self, libtype, solidx):
         if self.indtype == torch.float8_e4m3fnuz:
             ref, _ = torch._scaled_mm(self.inp,
                                       self.weights.t(),
+                                      scale_a=ONE,
+                                      scale_b=ONE,
                                       out_dtype=self.outdtype)
         else:
             ref = F.linear(self.inp, self.weights)

gradlib/setup.py

@@ -125,11 +125,14 @@
                 extra_compile_args={
                     'cxx': [
                         '-O3',
+                        '-DLEGACY_HIPBLAS_DIRECT=ON',
                     ],
                     'nvcc': [
-                        '-O3', '-U__CUDA_NO_HALF_OPERATORS__',
+                        '-O3',
+                        '-U__CUDA_NO_HALF_OPERATORS__',
                         '-U__CUDA_NO_HALF_CONVERSIONS__',
-                        "-ftemplate-depth=1024"
+                        "-ftemplate-depth=1024",
+                        '-DLEGACY_HIPBLAS_DIRECT=ON',
                     ] + extra_args
                 }))
         ext_modules.append(
@@ -142,11 +145,14 @@
                 extra_compile_args={
                     'cxx': [
                         '-O3',
+                        '-DLEGACY_HIPBLAS_DIRECT=ON',
                     ],
                     'nvcc': [
-                        '-O3', '-U__CUDA_NO_HALF_OPERATORS__',
+                        '-O3',
+                        '-U__CUDA_NO_HALF_OPERATORS__',
                         '-U__CUDA_NO_HALF_CONVERSIONS__',
-                        "-ftemplate-depth=1024"
+                        "-ftemplate-depth=1024",
+                        '-DLEGACY_HIPBLAS_DIRECT=ON',
                     ] + extra_args
                 }))
 

vllm/model_executor/layers/quantization/utils/w8a8_utils.py

@@ -10,7 +10,7 @@
 # providing scaling factor for result. This value is created
 # as global value to avoid multiple tensor allocations, and
 # can be removed once pytorch fixes the bug.
-TORCH_SCALED_MM_SCALE_RESULT = torch.ones(1).cuda() if is_hip() else None
+TORCH_DEVICE_IDENTITY = torch.ones(1).cuda() if is_hip() else None
 
 
 def cutlass_fp8_supported() -> bool:
@@ -132,20 +132,17 @@ def apply_fp8_linear(
         per_tensor_weights = (weight_scale.numel() == 1)
         per_tensor_activations = (x_scale.numel() == 1)
 
+        global TORCH_DEVICE_IDENTITY
+        if TORCH_DEVICE_IDENTITY.device != weight.device:
+            TORCH_DEVICE_IDENTITY = TORCH_DEVICE_IDENTITY.to(weight.device)
         if per_tensor_weights and per_tensor_activations:
             # Fused GEMM_DQ
-            global TORCH_SCALED_MM_SCALE_RESULT
-            if TORCH_SCALED_MM_SCALE_RESULT.device != weight.device:
-                TORCH_SCALED_MM_SCALE_RESULT = TORCH_SCALED_MM_SCALE_RESULT.to(
-                    weight.device)
-            output = torch._scaled_mm(
-                qinput,
-                weight,
-                out_dtype=out_dtype,
-                scale_a=x_scale,
-                scale_b=weight_scale,
-                scale_result=TORCH_SCALED_MM_SCALE_RESULT,
-                bias=bias)
+            output = torch._scaled_mm(qinput,
+                                      weight,
+                                      out_dtype=out_dtype,
+                                      scale_a=x_scale,
+                                      scale_b=weight_scale,
+                                      bias=bias)
             # A fix for discrepancy in scaled_mm which returns tuple
             # for torch < 2.5 and a single value in torch >= 2.5
             if type(output) is tuple and len(output) == 2:
@@ -173,6 +170,8 @@ def apply_fp8_linear(
             # Output in fp32 to allow subsequent ops to happen in-place
             output, _ = torch._scaled_mm(qinput,
                                          weight,
+                                         scale_a=TORCH_DEVICE_IDENTITY,
+                                         scale_b=TORCH_DEVICE_IDENTITY,
                                          out_dtype=torch.float32)
             # Unpad (undo num_token_padding)
             output = torch.narrow(output, 0, 0, input.shape[0])