diff --git a/csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu b/csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu index 0b6df02c7e..316a7d3752 100644 --- a/csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu +++ b/csrc/libtorch_stable/quantization/w8a8/fp8/per_token_group_quant.cu @@ -304,17 +304,9 @@ __global__ void per_token_group_quant_8bit_packed_register_kernel( const int sf_k_idx = blockIdx.x * kGroupsPerBlockX + sf_k_local; const int mn_idx = blockIdx.y * kRowsPerBlock + row_local; -#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)) - asm volatile("griddepcontrol.wait;"); -#endif - if (mn_idx >= tma_aligned_mn) { -#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)) - asm volatile("griddepcontrol.launch_dependents;"); -#endif return; } - const bool is_valid_group = (mn_idx < mn) && (sf_k_idx < groups_per_row); // Load 16 input elements (32 B) into registers as two adjacent uint4 @@ -425,10 +417,6 @@ __global__ void per_token_group_quant_8bit_packed_register_kernel( static_cast(mn_idx) * groups_per_row * GROUP_SIZE + sf_k_idx * GROUP_SIZE + lane_id * VEC_SIZE; *reinterpret_cast(group_output) = packed_out; - -#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 900)) - asm volatile("griddepcontrol.launch_dependents;"); -#endif } // Public entry point: register-resident packed quant kernel. @@ -509,29 +497,20 @@ void per_token_group_quant_8bit_packed(const torch::stable::Tensor& input, #define LAUNCH_REG_KERNEL_INST(T, DST_DTYPE, KX, RY) \ do { \ - cudaLaunchConfig_t config = {}; \ - config.gridDim = dim3(static_cast(blocks_x), \ - static_cast(blocks_y)); \ - config.blockDim = dim3(num_threads); \ - config.dynamicSmemBytes = 0; \ - config.stream = stream; \ - cudaLaunchAttribute attrs[1]; \ - attrs[0].id = cudaLaunchAttributeProgrammaticStreamSerialization; \ - attrs[0].val.programmaticStreamSerializationAllowed = 1; \ - config.numAttrs = 1; \ - config.attrs = attrs; \ - cudaLaunchKernelEx( \ - &config, \ - per_token_group_quant_8bit_packed_register_kernel, \ - static_cast(input.data_ptr()), output_q.data_ptr(), \ - reinterpret_cast(output_s_packed.data_ptr()), \ - static_cast(padded_groups_per_row), \ - static_cast(groups_per_row), static_cast(mn), \ - static_cast(output_q_mn_extent), \ - static_cast(tma_aligned_mn), num_scale_elems, \ - static_cast(eps), static_cast(min_8bit), \ - static_cast(max_8bit)); \ + dim3 grid(static_cast(blocks_x), \ + static_cast(blocks_y)); \ + dim3 block(num_threads); \ + per_token_group_quant_8bit_packed_register_kernel \ + <<>>( \ + static_cast(input.data_ptr()), output_q.data_ptr(), \ + reinterpret_cast(output_s_packed.data_ptr()), \ + static_cast(padded_groups_per_row), \ + static_cast(groups_per_row), static_cast(mn), \ + static_cast(output_q_mn_extent), \ + static_cast(tma_aligned_mn), num_scale_elems, \ + static_cast(eps), static_cast(min_8bit), \ + static_cast(max_8bit)); \ } while (0) #define LAUNCH_REG_KERNEL(T, DST_DTYPE) \ diff --git a/vllm/models/deepseek_v4/common/ops/fused_inv_rope_fp8_quant.py b/vllm/models/deepseek_v4/common/ops/fused_inv_rope_fp8_quant.py index 000bb51b20..97fc0962c2 100644 --- a/vllm/models/deepseek_v4/common/ops/fused_inv_rope_fp8_quant.py +++ b/vllm/models/deepseek_v4/common/ops/fused_inv_rope_fp8_quant.py @@ -37,8 +37,6 @@ def _fused_inv_rope_fp8_quant_per_head( ROPE_START: tl.constexpr, HALF_ROPE: tl.constexpr, TMA_ALIGNED_SCALES: tl.constexpr, - USE_GDC: tl.constexpr, - launch_pdl: tl.constexpr, # triton metadata ): # int64: stride multiply overflows int32 past num_tokens=32768 (IMA). pid_token = tl.program_id(0).to(tl.int64) @@ -48,9 +46,7 @@ def _fused_inv_rope_fp8_quant_per_head( head_in_group = pid_gh % heads_per_group global_head = pid_gh qb_start = head_in_group * CHUNKS_PER_HEAD - if USE_GDC: - tl.extra.cuda.gdc_launch_dependents() - tl.extra.cuda.gdc_wait() + # Padding rows in the TMA-aligned scale buffer: fill with zero and skip quant. if pid_token >= num_tokens: if TMA_ALIGNED_SCALES: @@ -247,8 +243,11 @@ def _fused_inv_rope_fp8_quant_kernel_impl( (scale_inner * tma_aligned_T, 1, tma_aligned_T), ) grid = (tma_aligned_T, n_groups * heads_per_group) - use_gdc = current_platform.is_arch_support_pdl() - pdl_kwargs = {"launch_pdl": True} if use_gdc else {} + pdl_kwargs = ( + {} + if current_platform.is_rocm() or current_platform.is_xpu() + else {"launch_pdl": False} + ) _fused_inv_rope_fp8_quant_per_head[grid]( o, positions, @@ -271,7 +270,6 @@ def _fused_inv_rope_fp8_quant_kernel_impl( ROPE_START=rope_start, HALF_ROPE=half_rope, TMA_ALIGNED_SCALES=tma_aligned_scales, - USE_GDC=use_gdc, num_stages=1, **pdl_kwargs, num_warps=1, diff --git a/vllm/utils/deep_gemm.py b/vllm/utils/deep_gemm.py index 1ddc93ff5e..3c884aad6c 100644 --- a/vllm/utils/deep_gemm.py +++ b/vllm/utils/deep_gemm.py @@ -177,22 +177,6 @@ def _import_deep_gemm(): return None -def _apply_pdl(mod, enable: bool = True) -> None: - mod_name = getattr(mod, "__name__", str(mod)) - try: - set_pdl_fn = getattr(mod, "set_pdl", None) - if set_pdl_fn is None: - return - set_pdl_fn(enable) - logger.info_once( - "DeepGEMM PDL %s on %s.", - "enabled" if enable else "disabled", - mod_name, - ) - except Exception as e: # noqa: BLE001 - logger.warning_once("Failed to set DeepGEMM PDL on %s: %s", mod_name, e) - - def _lazy_init() -> None: """Import deep_gemm and resolve symbols on first use.""" global _cublaslt_gemm_nt_impl @@ -235,9 +219,6 @@ def _lazy_init() -> None: if _dg is None: return - # Enable PDL for DeepGEMM on architectures that support it (SM90+). - if current_platform.is_arch_support_pdl(): - _apply_pdl(_dg, True) _cublaslt_gemm_nt_impl = getattr(_dg, "cublaslt_gemm_nt", None) _fp8_gemm_nt_impl = getattr(_dg, "fp8_gemm_nt", None) _fp8_einsum_impl = getattr(_dg, "fp8_einsum", None)