Update vllm kernels for ROCM (#2826)

* (vllm) updated vllm rocm kernels

* revert silu

* update partition size

* remove grouped_topk

* (nit) remove log

* update moe-kernels commit

Dockerfile_amd

@@ -234,6 +234,7 @@ FROM kernel-builder AS vllm-builder
 WORKDIR /usr/src
 
 COPY server/Makefile-vllm Makefile
+RUN pip install setuptools_scm
 
 # Build specific version of vllm
 RUN make build-vllm-rocm
@@ -267,6 +268,15 @@ COPY server/exllamav2_kernels/ .
 
 RUN python setup.py build
 
+FROM kernel-builder AS moe-kernels
+WORKDIR /usr/src
+ENV MOE_KERNELS_BRANCH=a67b35841774b2056a73806c36661134b5054edd
+ENV VLLM_TARGET_DEVICE=rocm
+RUN git clone https://github.com/danieldk/moe-kernels.git && \
+    cd moe-kernels && \
+    git checkout ${MOE_KERNELS_BRANCH} && \
+    python setup.py install
+
 FROM install_deps AS base-copy
 
 # Text Generation Inference base env
@@ -289,6 +299,9 @@ COPY --from=exllama-kernels-builder /usr/src/build/lib.linux-x86_64-cpython-311
 # Copy build artifacts from exllamav2 kernels builder
 COPY --from=exllamav2-kernels-builder /usr/src/build/lib.linux-x86_64-cpython-311 /opt/conda/lib/python3.11/site-packages
 
+# Copy build artifacts from moe kernels
+COPY --from=moe-kernels /usr/src/moe-kernels/build/lib.linux-x86_64-cpython-311 /opt/conda/lib/python3.11/site-packages
+
 # Install server
 COPY proto proto
 COPY server server

server/Makefile-vllm

@@ -1,4 +1,4 @@
-commit_rocm := 4e0929e6e4fa0a3d09d358715c288020ea9dc247
+commit_rocm := de990cd12537f78f74e40b5c8ee1a62d63d734dd
 
 build-vllm-rocm:
 	if [ ! -d 'vllm' ]; then \

server/text_generation_server/layers/attention/kv_cache.py

@@ -215,7 +215,9 @@ def paged_reshape_and_cache(
             raise ImportError(
                 f"Could not import vllm paged attention. Make sure your installation is correct. Complete error: {e}"
             )
-        ops.reshape_and_cache(key, value, key_cache, value_cache, slots, "auto", 1.0)
+        ops.reshape_and_cache(
+            key, value, key_cache, value_cache, slots, "auto", 1.0, 1.0
+        )
     elif SYSTEM == "ipex":
         import intel_extension_for_pytorch as ipex
 

server/text_generation_server/layers/attention/rocm.py

@@ -6,26 +6,42 @@
 from text_generation_server.layers.attention import Seqlen
 from text_generation_server.utils.log import log_master
 from loguru import logger
+import vllm._custom_ops as ops
 
 major, minor = torch.cuda.get_device_capability()
 is_sm75 = major == 7 and minor == 5
 
-_PARTITION_SIZE_V1V2 = 512
+_PARTITION_SIZE_V1V2 = 1024
 _PARTITION_SIZE_CUSTOM = 256
 
+_GPU_ARCH = torch.cuda.get_device_properties("cuda").gcnArchName
+_ON_MI250_MI300 = any(
+    arch in _GPU_ARCH for arch in ["gfx90a", "gfx940", "gfx941", "gfx942"]
+)
+
 use_triton = os.getenv("ROCM_USE_FLASH_ATTN_V2_TRITON", "").lower() in {"true", "1"}
 ENGINE = "triton" if use_triton else "ck"
 
 use_rocm_custom_paged_attn = os.getenv("ROCM_USE_CUSTOM_PAGED_ATTN", "1") != "0"
-try:
-    if use_rocm_custom_paged_attn:
-        from vllm._custom_C import paged_attention_custom
-except ImportError as e:
-    log_master(
-        logger.info,
-        f"Custom Paged Attention not available. Complete error: {e}",
+
+
+def _use_rocm_custom_paged_attention(
+    qtype: torch.dtype,
+    head_size: int,
+    block_size: int,
+    gqa_ratio: int,
+    max_seq_len: int,
+) -> bool:
+    # rocm custom page attention not support on navi (gfx1*)
+    return (
+        use_rocm_custom_paged_attn
+        and _ON_MI250_MI300
+        and (qtype == torch.half or qtype == torch.bfloat16)
+        and (head_size == 64 or head_size == 128)
+        and (block_size == 16 or block_size == 32)
+        and (gqa_ratio >= 1 and gqa_ratio <= 16)
+        and max_seq_len <= 131072
     )
-    use_rocm_custom_paged_attn = False
 
 
 def paged_attention(
@@ -66,13 +82,8 @@ def paged_attention(
 
     num_kv_heads = kv_cache.key.shape[1]
     gqa_ratio = num_heads // num_kv_heads
-    use_custom = (
-        use_rocm_custom_paged_attn
-        and (query.dtype == torch.half or query.dtype == torch.bfloat16)
-        and (head_size == 128 or head_size == 64)
-        and (block_size == 16 or block_size == 32)
-        and (gqa_ratio >= 1 and gqa_ratio <= 16)
-        and max_s <= 32768
+    use_custom = _use_rocm_custom_paged_attention(
+        query.dtype, head_size, block_size, gqa_ratio, max_s
     )
 
     if not use_custom:
@@ -90,8 +101,6 @@ def paged_attention(
     # V1 to avoid the overhead of reduction. Also, if the number of
     # sequences or heads is large, we use V1 since there is enough work
     # to parallelize.
-    import vllm._custom_ops as ops
-
     use_v1 = (
         max_s <= 8192
         and (max_num_partitions == 1 or num_seqs * num_heads > 512)
@@ -103,7 +112,7 @@ def paged_attention(
             query,
             kv_cache.key,
             kv_cache.value,
-            kv_head_mapping,
+            num_kv_heads,
             softmax_scale,
             block_tables,
             input_lengths,
@@ -112,6 +121,7 @@ def paged_attention(
             None,
             "auto",
             1.0,
+            1.0,
         )
     else:
         # Run PagedAttention V2.
@@ -137,7 +147,7 @@ def paged_attention(
                 query,
                 kv_cache.key,
                 kv_cache.value,
-                kv_head_mapping,
+                num_kv_heads,
                 softmax_scale,
                 block_tables,
                 input_lengths,
@@ -146,9 +156,10 @@ def paged_attention(
                 None,
                 "auto",
                 1.0,
+                1.0,
             )
         else:
-            paged_attention_custom(
+            ops.paged_attention_rocm(
                 out,
                 exp_sums,
                 max_logits,
@@ -164,6 +175,10 @@ def paged_attention(
                 max_s,
                 None,
                 "auto",
+                1.0,
+                1.0,
+                None,
+                _PARTITION_SIZE,
             )
 
     return out

server/text_generation_server/layers/layernorm.py

@@ -72,7 +72,7 @@ def forward(self, hidden_states, residual=None):
                 return normed_hidden_states, residual
 
 elif SYSTEM == "rocm":
-    from vllm._C import ops
+    import vllm._custom_ops as ops
 
     class FastLayerNorm(nn.LayerNorm):
         def forward(self, hidden_states, residual=None):
@@ -121,6 +121,27 @@ def forward(self, hidden_states, residual=None):
                 residual is not None,
             )
             return out, residual if residual is not None else hidden_states
+        elif SYSTEM == "rocm":
+            # We use VLLM RMSNorm kernel that can be compiled for RoCm, instead of Flash Attention ones that can not.
+            if residual is not None:
+                ops.fused_add_rms_norm(
+                    hidden_states,
+                    residual,
+                    self.weight.data,
+                    self.variance_epsilon,
+                )
+                return hidden_states, residual
+
+            residual = hidden_states
+
+            out = torch.empty_like(hidden_states)
+            ops.rms_norm(
+                out,
+                hidden_states,
+                self.weight.data,
+                self.variance_epsilon,
+            )
+            return out, residual
         elif hidden_states.shape[-1] > 8192:
             if residual is not None:
                 hidden_states += residual
@@ -164,20 +185,6 @@ def forward(self, hidden_states, residual=None):
                 res = hidden_states
 
             return normed_hidden_states, res
-        elif SYSTEM == "rocm":
-            # We use VLLM RMSNorm kernel that can be compiled for RoCm, instead of Flash Attention ones that can not.
-            if residual is not None:
-                hidden_states += residual
-            residual = hidden_states
-
-            out = torch.empty_like(hidden_states)
-            ops.rms_norm(
-                out,
-                hidden_states,
-                self.weight.data,
-                self.variance_epsilon,
-            )
-            return out, residual
         else:
             raise ValueError(
                 "Your system seem to be not supported. Please check your install or open an issue at https://github.com/huggingface/text-generation-inference/issues with a clear reproduction."

server/text_generation_server/layers/linear.py

@@ -11,10 +11,10 @@
 
     if ROCM_USE_SKINNY_GEMM:
         try:
-            from vllm import _custom_C
+            import vllm._custom_ops as ops
         except Exception as e:
             raise ImportError(
-                f"Could not load `vllm._custom_C` for ROCm skinny gemm. Full error: {e}"
+                f"Could not load `vllm._custom_ops` for ROCm skinny gemm. Full error: {e}"
             )
 
 
@@ -95,12 +95,12 @@ def forward(self, inp: torch.Tensor) -> torch.Tensor:
                 out = torch.empty(
                     inp_shape[0], weight.shape[0], dtype=inp.dtype, device=weight.device
                 )
-                _custom_C.wvSpltK(weight, inp, out, n, self.cu_count)
+                ops.wvSpltK(weight, inp, out, n, self.cu_count)
             elif m % 4 == 0 and n == 1 and k <= 8192:
                 out = torch.empty(
                     inp_shape[0], weight.shape[0], dtype=inp.dtype, device=weight.device
                 )
-                _custom_C.LLMM1(weight, inp, out, 4)
+                ops.LLMM1(weight, inp, out, 4)
             else:
                 out = F.linear(inp, weight)
 

server/text_generation_server/layers/moe/__init__.py

@@ -24,10 +24,7 @@
     UnquantizedWeight,
 )
 
-if SYSTEM == "rocm":
-    from .fused_moe_rocm import grouped_topk
-    from vllm.model_executor.layers.fused_moe import fused_topk
-elif SYSTEM == "ipex":
+if SYSTEM == "ipex":
     from intel_extension_for_pytorch.llm.modules import GatedMLPMOE
 else:
     from moe_kernels.fused_moe import fused_topk, grouped_topk

server/text_generation_server/layers/moe/unquantized.py

@@ -6,9 +6,7 @@
 from text_generation_server.utils.import_utils import SYSTEM
 from text_generation_server.utils.weights import UnquantizedWeight, Weights
 
-if SYSTEM == "rocm":
-    from vllm.model_executor.layers.fused_moe import fused_moe
-elif SYSTEM == "ipex":
+if SYSTEM == "ipex":
     from intel_extension_for_pytorch.llm.modules import GatedMLPMOE
 else:
     from moe_kernels.fused_moe import fused_moe

server/text_generation_server/layers/rotary.py

@@ -7,7 +7,7 @@
 if SYSTEM == "cuda":
     import rotary_emb
 elif SYSTEM == "rocm":
-    from vllm._C import ops
+    import vllm._custom_ops as ops
 elif SYSTEM == "ipex":
     import intel_extension_for_pytorch as ipex
 

server/text_generation_server/models/custom_modeling/flash_cohere_modeling.py

@@ -75,7 +75,7 @@ def forward(
 
             rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
         elif SYSTEM == "rocm":
-            from vllm._C import ops
+            import vllm._custom_ops as ops
 
             # NOTE: On RoCm systems, we use a ROPE implementatation adapted from VLLM which launches a single kernel for both query/key, contrary to flash-attn implementation used on NVIDIA systems.
             # Compiling flash-attn rotary on RoCm, it appears hipcc is unable to unroll loops, resulting in an even slower inference compared to eager: https://github.com/pytorch/pytorch/issues/113773

server/text_generation_server/models/custom_modeling/flash_dbrx_modeling.py

@@ -23,9 +23,7 @@
 from text_generation_server.layers.attention.kv_cache import get_kv_scales
 from text_generation_server.utils.import_utils import SYSTEM
 
-if SYSTEM == "rocm":
-    from vllm.model_executor.layers.fused_moe import fused_moe
-elif SYSTEM == "ipex":
+if SYSTEM == "ipex":
     from intel_extension_for_pytorch.llm.modules import GatedMLPMOE
 else:
     from moe_kernels.fused_moe import fused_moe

server/text_generation_server/models/custom_modeling/flash_deepseek_v2_modeling.py

@@ -43,9 +43,9 @@
 
 if SYSTEM == "rocm":
     try:
-        from vllm import _custom_C
+        import vllm._custom_ops as ops
     except Exception as e:
-        raise ImportError(f"Could not load `vllm._custom_C`. Full error: {e}")
+        raise ImportError(f"Could not load `vllm._custom_ops`. Full error: {e}")
 
 
 class DeepseekV2Config(PretrainedConfig):
@@ -408,7 +408,7 @@ def forward(self, hidden_states: torch.Tensor, reduce: bool = True):
                 dtype=hidden_states.dtype,
                 device="cuda",
             )
-            _custom_C.LLMM_Silu(self.gate_up_proj.linear.weight, hidden_states, out, 8)
+            ops.LLMM_Silu(self.gate_up_proj.linear.weight, hidden_states, out, 8)
             return self.down_proj(out, reduce=reduce)
         else:
             gate_up_states = self.gate_up_proj(hidden_states)

server/text_generation_server/models/custom_modeling/flash_gptj_modeling.py

@@ -91,7 +91,7 @@ def forward(
 
             rotary_emb.apply_rotary(k1, k2, cos, sin, k1, k2, False)
         elif SYSTEM == "rocm":
-            from vllm._C import ops
+            import vllm._custom_ops as ops
 
             # NOTE: On RoCm systems, we use a ROPE implementatation adapted from VLLM which launches a single kernel for both query/key, contrary to flash-attn implementation used on NVIDIA systems.
             # Compiling flash-attn rotary on RoCm, it appears hipcc is unable to unroll loops, resulting in an even slower inference compared to eager: https://github.com/pytorch/pytorch/issues/113773