SM90 (Hopper) FP4 MegaMoE fused kernel with swapAB small-batch path

csrc/apis/sm90_mega.hpp

@@ -1,17 +1,22 @@
 #pragma once
 
 #include <functional>
+#include <limits>
+#include <optional>
+#include <string>
+#include <tuple>
+#include <vector>
 
 #include "mega.hpp"
+#include "../jit/device_runtime.hpp"
+#include "../jit_kernels/impls/sm90_fp8_fp4_mega_moe.hpp"
 #include "../jit_kernels/impls/sm90_fp8_mega_moe.hpp"
 #include "../jit_kernels/impls/sm90_mega_moe_pre_dispatch.hpp"
+#include "../utils/layout.hpp"
+#include "../utils/system.hpp"
 
 namespace deep_gemm::mega {
 
-static int get_token_alignment_for_sm90_mega_moe() {
-    return layout::kLCMCandidateBlockM;
-}
-
 static void mega_moe_pre_dispatch_sm90(
     const torch::Tensor& x,
     const torch::Tensor& topk_idx,
@@ -30,6 +35,68 @@ static void mega_moe_pre_dispatch_sm90(
         num_tokens, group_size, routed_scaling_factor);
 }
 
+static bool is_packed_fp4_storage_sm90(const torch::Tensor& t) {
+    return t.scalar_type() == kPackedFP4 or t.scalar_type() == torch::kByte;
+}
+
+static std::tuple<int, int, int> check_grouped_ab_sm90_fp4_mega_moe(const torch::Tensor& ab) {
+    const auto [num_groups, mn, packed_k] = get_shape<3>(ab);
+    DG_HOST_ASSERT(is_packed_fp4_storage_sm90(ab));
+    DG_HOST_ASSERT(get_major_type_ab(ab) == cute::UMMA::Major::K);
+    DG_HOST_ASSERT(packed_k > 0 and packed_k % 64 == 0);
+    return {num_groups, mn, packed_k * 2};
+}
+
+static void check_sm90_fp4_sfb_layout(const torch::Tensor& sf,
+                                      const int& mn, const int& k,
+                                      const int& num_groups) {
+    DG_HOST_ASSERT(sf.scalar_type() == torch::kInt);
+    DG_HOST_ASSERT(sf.dim() == 3);
+    DG_HOST_ASSERT(sf.size(0) == num_groups);
+    DG_HOST_ASSERT(sf.size(1) == mn);
+    DG_HOST_ASSERT(sf.size(2) == ceil_div(k, 128));
+    DG_HOST_ASSERT(sf.is_contiguous());
+}
+
+struct FP4SM90APIDefaults {
+    bool wide_load_decode;
+    bool early_b_decode;
+    bool decode_done_mbarrier;
+    bool ss_nsplit;
+    bool swap_ab;
+};
+
+static FP4SM90APIDefaults get_fp4_sm90_api_defaults(
+    const int& num_experts_per_rank, const int& num_tokens, const int& num_topk,
+    const bool& use_situ) {
+    // Select the decode, prefill, and swapAB features from the same routing
+    // density so the generated kernel uses a coherent configuration bundle.
+    const float expected_tokens_per_expert =
+        static_cast<float>(num_tokens) * num_topk / num_experts_per_rank;
+    // SiTU uses the analytic cost model while SwiGLU keeps its scalar boundary;
+    // share the predicate with the block-config heuristics so the feature
+    // bundle never mixes.
+    const bool prefill_band = is_fp4_sm90_prefill_band(
+        expected_tokens_per_expert, use_situ);
+    const bool decode_band =
+        expected_tokens_per_expert > 0.0f and !prefill_band;
+    // SwiGLU uses a fixed low-density swapAB boundary. SiTU derives swapAB
+    // from the same cost comparison as its decode/prefill decision.
+    const bool swap_ab =
+        use_situ
+            ? (get_fp4_sm90_situ_config_kind(expected_tokens_per_expert) ==
+               FP4SM90ConfigKind::kSwapAB)
+            : (decode_band and
+               expected_tokens_per_expert < kSM90FP4SwiGLUSwapABMaxE);
+    return {
+        /*wide_load_decode=*/ decode_band,
+        /*early_b_decode=*/ prefill_band,
+        /*decode_done_mbarrier=*/ expected_tokens_per_expert > 0.0f,
+        /*ss_nsplit=*/ prefill_band,
+        /*swap_ab=*/ swap_ab
+    };
+}
+
 static std::tuple<int64_t, std::function<std::tuple<torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor, torch::Tensor>(const torch::Tensor&)>>
 get_symm_buffer_size_for_sm90_mega_moe(
     const int& num_ranks, const int& num_experts,
@@ -38,16 +105,17 @@ get_symm_buffer_size_for_sm90_mega_moe(
     const bool& use_fp8_dispatch, const std::string& activation) {
     DG_HOST_ASSERT(num_experts % num_ranks == 0);
     DG_HOST_ASSERT(use_fp8_dispatch);
-    DG_HOST_ASSERT(activation == "swiglu");
-
+    DG_HOST_ASSERT(activation == "swiglu" or activation == "situ");
     const auto workspace = layout::SM90Workspace(
         nullptr, num_ranks, num_experts, num_max_tokens_per_rank, num_topk);
 
     const auto fp8_token_layout = layout::Data(hidden);
     const auto bf16_token_layout = layout::Data(hidden * 2);
     const auto fp8_intermediate_token_layout = layout::Data(intermediate_hidden);
-    const auto fp8_sf_layout = layout::Data(hidden / 32);
-    const auto fp8_intermediate_sf_layout = layout::Data(intermediate_hidden / 16);
+    const auto fp8_sf_layout =
+        layout::Data(hidden * static_cast<int>(sizeof(float)) / kSM90FP4L1ActSFGranK);
+    const auto fp8_intermediate_sf_layout =
+        layout::Data(intermediate_hidden * static_cast<int>(sizeof(float)) / kSM90FP4L2ActSFGranK);
     const auto input_topk_idx_layout = layout::Data(num_topk * sizeof(int64_t), false);
     const auto input_topk_weights_layout = layout::Data(num_topk * sizeof(float), false);
     const auto l1_topk_weights_layout = layout::Data(sizeof(float), false);
@@ -66,13 +134,10 @@ get_symm_buffer_size_for_sm90_mega_moe(
         input_topk_idx_buffer.get_end_ptr());
 
     const auto num_max_pool_tokens = static_cast<int>(workspace.num_max_pool_tokens);
-    int num_max_padded_sf_pool_tokens = 0;
-    for (int block_m: layout::kCandidateBlockM) {
-        num_max_padded_sf_pool_tokens = std::max(
-            num_max_padded_sf_pool_tokens,
-            layout::get_num_sf_ring_tokens(num_max_pool_tokens, block_m)
-        );
-    }
+    constexpr int kMinSM90MegaMoEBlockM = 64;
+    const auto num_max_padded_sf_pool_tokens = static_cast<int>(
+        layout::get_num_sf_ring_tokens(
+            num_max_pool_tokens, kMinSM90MegaMoEBlockM));
 
     const auto l1_token_buffer = layout::Buffer(
         fp8_token_layout, 1, num_max_pool_tokens,
@@ -96,6 +161,8 @@ get_symm_buffer_size_for_sm90_mega_moe(
         l2_sf_buffer.get_end_ptr());
 
     DG_HOST_ASSERT(hidden % 128 == 0 and intermediate_hidden % 128 == 0);
+    DG_HOST_ASSERT(hidden % kSM90FP4L1ActSFGranK == 0);
+    DG_HOST_ASSERT(intermediate_hidden % kSM90FP4L2ActSFGranK == 0);
 
     auto slice_input_buffers = [=](const torch::Tensor& buffer) {
         auto x = torch::from_blob(
@@ -104,7 +171,7 @@ get_symm_buffer_size_for_sm90_mega_moe(
             torch::TensorOptions().dtype(torch::kFloat8_e4m3fn).device(buffer.device()));
         auto x_sf = torch::from_blob(
             math::advance_ptr(buffer.data_ptr(), reinterpret_cast<int64_t>(input_sf_buffer.base)),
-            {num_max_tokens_per_rank, hidden / 128},
+            {num_max_tokens_per_rank, hidden / kSM90FP4L1ActSFGranK},
             torch::TensorOptions().dtype(torch::kFloat32).device(buffer.device()));
         auto topk_idx = torch::from_blob(
             math::advance_ptr(buffer.data_ptr(), reinterpret_cast<int64_t>(input_topk_idx_buffer.base)),
@@ -120,7 +187,7 @@ get_symm_buffer_size_for_sm90_mega_moe(
             torch::TensorOptions().dtype(torch::kFloat8_e4m3fn).device(buffer.device()));
         auto l1_acts_sf = torch::from_blob(
             math::advance_ptr(buffer.data_ptr(), reinterpret_cast<int64_t>(l1_sf_buffer.base)),
-            {num_max_padded_sf_pool_tokens, hidden / 128},
+            {num_max_padded_sf_pool_tokens, hidden / kSM90FP4L1ActSFGranK},
             {1, num_max_padded_sf_pool_tokens},
             torch::TensorOptions().dtype(torch::kFloat32).device(buffer.device()));
         auto l2_acts = torch::from_blob(
@@ -129,7 +196,7 @@ get_symm_buffer_size_for_sm90_mega_moe(
             torch::TensorOptions().dtype(torch::kFloat8_e4m3fn).device(buffer.device()));
         auto l2_acts_sf = torch::from_blob(
             math::advance_ptr(buffer.data_ptr(), reinterpret_cast<int64_t>(l2_sf_buffer.base)),
-            {num_max_padded_sf_pool_tokens, intermediate_hidden / 64},
+            {num_max_padded_sf_pool_tokens, intermediate_hidden / kSM90FP4L2ActSFGranK},
             {1, num_max_padded_sf_pool_tokens},
             torch::TensorOptions().dtype(torch::kFloat32).device(buffer.device()));
         return std::make_tuple(x, x_sf, topk_idx, topk_weights, l1_acts, l1_acts_sf, l2_acts, l2_acts_sf);
@@ -221,4 +288,110 @@ static void fp8_mega_moe(
         sym_buffer.zero_();
 }
 
+static void fp8_fp4_mega_moe_sm90(
+    const torch::Tensor& y,
+    const std::tuple<torch::Tensor, torch::Tensor>& l1_weights_tuple,
+    const std::tuple<torch::Tensor, torch::Tensor>& l2_weights_tuple,
+    const std::optional<torch::Tensor>& cumulative_local_expert_recv_stats,
+    const torch::Tensor& sym_buffer,
+    const std::vector<int64_t>& sym_buffer_ptrs, const int& rank_idx,
+    const int& num_max_tokens_per_rank,
+    const int& num_experts, const int& num_topk,
+    const std::tuple<int, int, int>& recipe,
+    const std::string& activation,
+    const std::optional<float>& activation_clamp_opt,
+    const std::optional<float>& activation_alpha_opt,
+    const std::optional<float>& activation_linear_beta_opt,
+    const bool& fast_math
+) {
+    const auto [l1_weights, l1_weights_sf] = l1_weights_tuple;
+    const auto [l2_weights, l2_weights_sf] = l2_weights_tuple;
+
+    const auto arch_major = device_runtime->get_arch_major();
+    DG_HOST_ASSERT(arch_major == 9);
+
+    const auto num_tokens = static_cast<int>(y.size(0));
+    const auto [rm, rn, rk] = recipe;
+    DG_HOST_ASSERT(rm == 1 and rn == 1 and rk == 32);
+    DG_HOST_ASSERT(activation == "swiglu" or activation == "situ");
+    const bool use_situ = activation == "situ";
+    DG_HOST_ASSERT(not use_situ or not activation_clamp_opt.has_value());
+
+    const auto activation_clamp =
+        activation_clamp_opt.value_or(std::numeric_limits<float>::infinity());
+    DG_HOST_ASSERT(activation_clamp >= 0);
+    const auto activation_alpha = activation_alpha_opt.value_or(4.0f);
+    const auto activation_linear_beta = activation_linear_beta_opt.value_or(25.0f);
+    DG_HOST_ASSERT(not use_situ or
+                   (activation_alpha > 0.0f and activation_linear_beta > 0.0f));
+
+    const auto [num_experts_per_rank, intermediate_hidden_2, hidden] =
+        check_grouped_ab_sm90_fp4_mega_moe(l1_weights);
+    const auto [num_experts_per_rank_, hidden_, intermediate_hidden] =
+        check_grouped_ab_sm90_fp4_mega_moe(l2_weights);
+    DG_HOST_ASSERT(num_tokens <= num_max_tokens_per_rank);
+    DG_HOST_ASSERT(num_experts_per_rank == num_experts_per_rank_);
+    DG_HOST_ASSERT(hidden == hidden_);
+    DG_HOST_ASSERT(intermediate_hidden_2 == 2 * intermediate_hidden);
+    DG_HOST_ASSERT(l1_weights.is_contiguous() and l2_weights.is_contiguous());
+    DG_HOST_ASSERT(hidden % 128 == 0 and intermediate_hidden % 128 == 0);
+    DG_HOST_ASSERT(intermediate_hidden / 64 <= 64);
+
+    check_sm90_fp4_sfb_layout(l1_weights_sf, intermediate_hidden * 2, hidden,
+                              num_experts_per_rank);
+    check_sm90_fp4_sfb_layout(l2_weights_sf, hidden, intermediate_hidden,
+                              num_experts_per_rank);
+
+    if (cumulative_local_expert_recv_stats.has_value()) {
+        DG_HOST_ASSERT(cumulative_local_expert_recv_stats->scalar_type() == torch::kInt);
+        DG_HOST_ASSERT(cumulative_local_expert_recv_stats->numel() == num_experts_per_rank);
+        DG_HOST_ASSERT(cumulative_local_expert_recv_stats->is_contiguous());
+    }
+
+    const auto num_ranks = static_cast<int>(sym_buffer_ptrs.size());
+    const auto num_experts_ = num_experts_per_rank * num_ranks;
+    const auto [num_required_bytes, slice] = get_symm_buffer_size_for_sm90_mega_moe(
+        num_ranks, num_experts,
+        num_max_tokens_per_rank, num_topk,
+        hidden, intermediate_hidden,
+        true, activation);
+    DG_HOST_ASSERT(sym_buffer.nbytes() >= static_cast<size_t>(num_required_bytes));
+    DG_HOST_ASSERT(num_experts == num_experts_);
+
+    const auto [x, x_sf, topk_idx, topk_weights, l1_acts, l1_acts_sf, l2_acts, l2_acts_sf] = slice(sym_buffer);
+    (void)x;
+    (void)x_sf;
+    (void)topk_idx;
+    (void)topk_weights;
+
+    DG_HOST_ASSERT(get_env<int>("DG_USE_FP4_ACTS") == 0);
+    DG_HOST_ASSERT(get_env<int>("DG_USE_FP8_COMBINE") == 0);
+
+    const auto fp4_defaults = get_fp4_sm90_api_defaults(
+        num_experts_per_rank, num_tokens, num_topk, use_situ);
+    sm90_fp8_fp4_mega_moe(y,
+                          l1_acts, l1_acts_sf,
+                          l2_acts, l2_acts_sf,
+                          l1_weights, l2_weights,
+                          l1_weights_sf, l2_weights_sf,
+                          cumulative_local_expert_recv_stats,
+                          sym_buffer_ptrs,
+                          rank_idx, num_max_tokens_per_rank,
+                          num_experts_per_rank,
+                          num_tokens, num_topk,
+                          hidden, intermediate_hidden,
+                          activation,
+                          activation_alpha,
+                          activation_linear_beta,
+                          activation_clamp, fast_math,
+                          fp4_defaults.wide_load_decode,
+                          fp4_defaults.early_b_decode,
+                          fp4_defaults.decode_done_mbarrier,
+                          fp4_defaults.ss_nsplit,
+                          fp4_defaults.swap_ab);
+
+    if (get_env<int>("DG_COMM_KERNEL_DEBUG"))
+        sym_buffer.zero_();
+}
+
 } // namespace deep_gemm::mega