aiconfigurator_core/engine/

spec.rs

// SPDX-FileCopyrightText: Copyright (c) 2025-2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
// SPDX-License-Identifier: Apache-2.0

//! `EngineSpec`: the serializable engine wire format.
//!
//! `EngineSpec` is what Python's `compile_engine` emits and what the
//! Rust `Engine` consumes. It bundles the engine identity
//! ([`EngineConfig`], needed later to load the matching [`PerfDatabase`]) with
//! the precompiled context / generation op lists. Each op is an
//! [`OpSpec`] — a public alias for the crate's [`Op`] enum — and the lists
//! round-trip through bincode, including the recursive `Overlap` / `Fallback`
//! children.
//!
//! ## Vision is never on the wire
//!
//! [`Op::Vision`] derives serde with every other variant (it remains part of
//! the shared session path), but a compiled `EngineSpec` never
//! contains a `Vision` op: `compile_engine` decomposes the vision encoder
//! into its child `Gemm` / `EncoderAttention` / `Elementwise` ops, each an
//! existing variant. The type round-trips soundly (see the test below); the
//! constraint is purely a producer-side rule, not enforced by the enum.
//!
//! [`PerfDatabase`]: crate::perf_database::PerfDatabase

use serde::{Deserialize, Serialize};

use crate::common::error::AicError;
use crate::{EngineConfig, ENGINE_SPEC_SCHEMA_VERSION};

/// Public name for the serializable op. Aliases the crate's [`Op`] enum so
/// the "OpSpec" surface exists without duplicating the definition.
pub use crate::operators::op::Op as OpSpec;

/// Serializable compiled engine.
///
/// `schema_version` guards forward/backward compatibility; `engine` carries
/// the identity used to load the matching perf database; `context_ops` and
/// `generation_ops` are the precompiled op lists the runner iterates.
///
/// `EngineSpec` derives serde so JSON / other self-describing formats work
/// directly. The **bincode** wire format, however, must go through
/// [`EngineSpec::to_bincode`] / [`EngineSpec::from_bincode`], NOT
/// `bincode::serialize(&spec)` directly: [`EngineConfig`] uses
/// `#[serde(flatten)]` (load-bearing for the flat ctypes FFI contract), and
/// bincode 1.x cannot serialize a flattened struct (it emits a map of unknown
/// length → `SequenceMustHaveLength`). The helpers sidestep this by
/// JSON-encoding the `engine` field inside the bincode payload, keeping
/// `EngineConfig` the single source of truth (no mirror struct, no drift).
#[derive(Clone, Debug, PartialEq, Serialize, Deserialize)]
pub struct EngineSpec {
    pub schema_version: u32,
    /// Engine identity (model / system / backend / parallelism / quant).
    /// Needed to locate and load the `PerfDatabase`.
    pub engine: EngineConfig,
    /// Context-phase ops, in execution order. Never contains `OpSpec::Vision`
    /// (decomposed into child ops at compile time).
    pub context_ops: Vec<OpSpec>,
    /// Generation-phase ops, in execution order.
    pub generation_ops: Vec<OpSpec>,
}

/// Private bincode payload. `engine` is carried as a JSON string so the
/// `#[serde(flatten)]` on [`EngineConfig`] never reaches the bincode
/// serializer (which rejects unknown-length maps). The op lists are plain
/// `Vec<OpSpec>` (no flatten) and bincode-serialize directly.
#[derive(Serialize, Deserialize)]
struct BincodeWire {
    schema_version: u32,
    engine_json: String,
    context_ops: Vec<OpSpec>,
    generation_ops: Vec<OpSpec>,
}

impl EngineSpec {
    /// Build a spec, stamping the current [`ENGINE_SPEC_SCHEMA_VERSION`].
    pub fn new(
        engine: EngineConfig,
        context_ops: Vec<OpSpec>,
        generation_ops: Vec<OpSpec>,
    ) -> Self {
        Self {
            schema_version: ENGINE_SPEC_SCHEMA_VERSION,
            engine,
            context_ops,
            generation_ops,
        }
    }

    /// Serialize to the bincode wire format. The `engine` field is
    /// JSON-encoded inside the payload (see the struct docs) so bincode never
    /// sees `EngineConfig`'s flattened layout.
    pub fn to_bincode(&self) -> Result<Vec<u8>, AicError> {
        let engine_json = serde_json::to_string(&self.engine)
            .map_err(|e| AicError::EngineSpec(format!("engine JSON encode: {e}")))?;
        let wire = BincodeWire {
            schema_version: self.schema_version,
            engine_json,
            context_ops: self.context_ops.clone(),
            generation_ops: self.generation_ops.clone(),
        };
        bincode::serialize(&wire).map_err(|e| AicError::EngineSpec(format!("bincode encode: {e}")))
    }

    /// Deserialize from the bincode wire format produced by [`Self::to_bincode`].
    ///
    /// The `schema_version` prefix is read and validated **before** the
    /// variable-layout op payloads are decoded. bincode is not self-describing,
    /// so a producer/consumer op-layout skew (e.g. a newer producer that added
    /// serialized fields to an `OpSpec`) would otherwise fail deep inside the
    /// payload with a generic `bincode decode: io error`, masking the real
    /// cause. Reading the leading version first lets a version mismatch surface
    /// as a clear [`AicError::UnsupportedSchemaVersion`] instead.
    pub fn from_bincode(bytes: &[u8]) -> Result<Self, AicError> {
        // `schema_version` is the first field of `BincodeWire`, so it is the
        // first value in the byte stream. Decode just it and gate on it before
        // touching the op lists (which is where a layout skew would fail).
        let mut cursor = std::io::Cursor::new(bytes);
        let schema_version: u32 = bincode::deserialize_from(&mut cursor).map_err(|e| {
            AicError::EngineSpec(format!(
                "bincode decode of the leading schema_version prefix failed \
                 (payload is {} bytes; too short or not an EngineSpec wire buffer): {e}",
                bytes.len()
            ))
        })?;
        if schema_version != ENGINE_SPEC_SCHEMA_VERSION {
            return Err(AicError::UnsupportedSchemaVersion {
                kind: "EngineSpec",
                got: schema_version,
                expected: ENGINE_SPEC_SCHEMA_VERSION,
            });
        }
        let wire: BincodeWire = bincode::deserialize(bytes).map_err(|e| {
            AicError::EngineSpec(format!(
                "bincode decode of the op payloads failed at matching \
                 schema_version {schema_version} — this indicates op-layout drift \
                 within the same version (an OpSpec changed without a \
                 ENGINE_SPEC_SCHEMA_VERSION bump) or a corrupt payload, not a \
                 version skew: {e}"
            ))
        })?;
        let engine: EngineConfig = serde_json::from_str(&wire.engine_json).map_err(|e| {
            AicError::EngineSpec(format!(
                "engine JSON decode failed at schema_version {schema_version}: {e}"
            ))
        })?;
        Ok(Self {
            schema_version: wire.schema_version,
            engine,
            context_ops: wire.context_ops,
            generation_ops: wire.generation_ops,
        })
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use std::collections::BTreeMap;

    use crate::common::enums::{
        BackendKind, CommQuantMode, FmhaQuantMode, GemmQuantMode, KvCacheQuantMode, MoeQuantMode,
    };
    use crate::operators::moe_dispatch::DispatchFlavor;
    use crate::operators::op::{FallbackOp, OverlapOp};
    use crate::operators::{
        ContextAttentionOp, ContextMlaOp, CustomAllReduceOp, DsaModuleOp, Dsv4MegaMoeOp,
        Dsv4ModuleOp, ElementwiseOp, EmbeddingOp, EncoderAttentionOp, GdnOp, GemmOp,
        GenerationAttentionOp, GenerationMlaOp, KdaOp, Mamba2Op, MhcModuleOp, MlaBmmOp,
        MlaModuleOp, MoEDispatchOp, MoeAllToAllOp, MoeExpertComputeOp, MoeOp, NcclOp, P2POp,
        VisionEncoderOp, WideEpContextMlaOp, WideEpGenerationMlaOp,
    };
    use crate::perf_database::dsv4::AttnKind;
    use crate::{
        DataType, ParallelMapping, QuantizationConfig, SpeculativeConfig,
        ENGINE_CONFIG_SCHEMA_VERSION,
    };

    // ---- Representative-value builders for each Op variant ----

    fn gemm() -> GemmOp {
        GemmOp {
            name: "qkv_gemm".into(),
            scale_factor: 2.0,
            n: 4096,
            k: 4096,
            quant_mode: GemmQuantMode::Fp8,
            scale_num_tokens: 0,
            low_precision_input: true,
            seq_split: 1,
            below_grid_sol: false,
        }
    }

    fn embedding() -> EmbeddingOp {
        EmbeddingOp {
            name: "embedding".into(),
            scale_factor: 1.0,
            vocab_size: 128_256,
            hidden_size: 4096,
            quant_mode: GemmQuantMode::Bfloat16,
            seq_split: 1,
        }
    }

    fn elementwise() -> ElementwiseOp {
        ElementwiseOp {
            name: "rmsnorm".into(),
            scale_factor: 1.5,
            bytes_per_token: 8192.0,
            scale_num_tokens: 1,
            seq_split: 1,
        }
    }

    fn context_attention() -> ContextAttentionOp {
        ContextAttentionOp {
            name: "context_attention".into(),
            scale_factor: 1.0,
            n: 32,
            n_kv: 8,
            head_size: 128,
            window_size: 0,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
            fmha_quant_mode: FmhaQuantMode::Bfloat16,
            use_qk_norm: true,
            cp_size: 1,
            // Multi-entry so the round-trip proves the whole Vec<String>
            // survives, not just a single-element degenerate case.
            lane_order: vec!["trtllm_mha".into(), "flashinfer".into(), "default".into()],
        }
    }

    fn generation_attention() -> GenerationAttentionOp {
        GenerationAttentionOp {
            name: "generation_attention".into(),
            scale_factor: 1.0,
            n: 32,
            n_kv: 8,
            head_size: 128,
            window_size: 4096,
            kv_cache_dtype: KvCacheQuantMode::Int8,
            lane_order: vec!["triton".into(), "trtllm_mha".into(), "default".into()],
        }
    }

    fn encoder_attention() -> EncoderAttentionOp {
        EncoderAttentionOp {
            name: "encoder_attention".into(),
            scale_factor: 1.0,
            n: 16,
            head_size: 80,
            fmha_quant_mode: FmhaQuantMode::Fp8,
            partial_rotary_factor: 0.0,
        }
    }

    fn context_mla() -> ContextMlaOp {
        ContextMlaOp {
            name: "context_mla".into(),
            scale_factor: 1.0,
            num_heads: 128,
            kv_cache_dtype: KvCacheQuantMode::Bfloat16,
            fmha_quant_mode: FmhaQuantMode::Bfloat16,
            cp_size: 1,
        }
    }

    fn generation_mla() -> GenerationMlaOp {
        GenerationMlaOp {
            name: "generation_mla".into(),
            scale_factor: 1.0,
            num_heads: 128,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
        }
    }

    fn mla_module() -> MlaModuleOp {
        MlaModuleOp {
            name: "context_mla_module".into(),
            scale_factor: 1.0,
            num_heads: 128,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
            fmha_quant_mode: FmhaQuantMode::Fp8,
            gemm_quant_mode: GemmQuantMode::Fp8Block,
            native_num_heads: Some(128),
        }
    }

    fn mla_bmm() -> MlaBmmOp {
        MlaBmmOp {
            name: "mla_bmm_pre".into(),
            scale_factor: 1.0,
            num_heads: 128,
            quant_mode: GemmQuantMode::Bfloat16,
            is_pre: true,
        }
    }

    fn moe() -> MoeOp {
        MoeOp {
            name: "moe".into(),
            scale_factor: 1.0,
            hidden_size: 7168,
            inter_size: 2048,
            topk: 8,
            num_experts: 256,
            moe_tp_size: 1,
            moe_ep_size: 8,
            attention_dp_size: 1,
            quant_mode: MoeQuantMode::Fp8Block,
            workload_distribution: "power_law_1.2".into(),
            is_gated: true,
            moe_backend: None,
            enable_eplb: false,
            is_context: false,
        }
    }

    fn moe_dispatch() -> MoEDispatchOp {
        MoEDispatchOp {
            name: "moe_dispatch".into(),
            scale_factor: 1.0,
            hidden_size: 7168,
            topk: 8,
            num_experts: 256,
            moe_tp_size: 1,
            moe_ep_size: 8,
            attention_dp_size: 8,
            pre_dispatch: true,
            backend: BackendKind::Trtllm,
            flavor: DispatchFlavor::TrtllmAlltoall,
            comm_quant: CommQuantMode::Half,
            moe_quant: MoeQuantMode::Fp8Block,
            attn_cp_size: 1,
            is_context: false,
            sms: 12,
            scale_num_tokens: 1,
            attn_ar_modeled: false,
        }
    }

    fn custom_all_reduce() -> CustomAllReduceOp {
        CustomAllReduceOp {
            name: "custom_all_reduce".into(),
            scale_factor: 1.0,
            hidden_size: 4096,
            tp_size: 8,
            quant: CommQuantMode::Half,
            seq_split: 1,
        }
    }

    fn nccl() -> NcclOp {
        NcclOp {
            name: "nccl_all_reduce".into(),
            scale_factor: 1.0,
            hidden_size: 4096.0,
            num_gpus: 8,
            dtype: CommQuantMode::Half,
            operation: "all_reduce".into(),
            seq_split: 1,
        }
    }

    fn p2p() -> P2POp {
        P2POp {
            name: "p2p".into(),
            scale_factor: 1.0,
            pp_size: 4,
            hidden_size: 4096,
            seq_split: 1,
        }
    }

    fn vision() -> VisionEncoderOp {
        VisionEncoderOp {
            name: "vision_encoder".into(),
            scale_factor: 1.0,
            num_layers: 24,
            num_heads: 16,
            head_size: 80,
            hidden_size: 1280,
            intermediate_size: 5120,
            fmha_quant: FmhaQuantMode::Bfloat16,
            gemm_quant: GemmQuantMode::Bfloat16,
        }
    }

    fn dsa_module() -> DsaModuleOp {
        DsaModuleOp {
            name: "dsa_module".into(),
            scale_factor: 1.0,
            num_heads: 128,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
            fmha_quant_mode: FmhaQuantMode::Fp8,
            gemm_quant_mode: GemmQuantMode::Fp8Block,
            architecture: "DeepseekV32ForCausalLM".into(),
            index_topk: 2048,
            cp_size: 1,
            full_frac: 1.0,
            attn_projection_quant_modes: None,
        }
    }

    fn msa_module() -> crate::operators::MsaModuleOp {
        crate::operators::MsaModuleOp {
            name: "context_attention".into(),
            scale_factor: 62.0,
            num_heads: 8,
            num_kv_heads: 1,
            hidden_size: 7168,
            head_dim: 128,
            v_head_dim: 128,
            index_n_heads: 64,
            index_head_dim: 128,
            index_topk: 2048,
            block_size: 64,
            kv_cache_dtype: KvCacheQuantMode::Bfloat16,
            fmha_quant_mode: FmhaQuantMode::Bfloat16,
            gemm_quant_mode: GemmQuantMode::Fp8Block,
            dsa_architecture: "GlmMoeDsaForCausalLM".into(),
            dsa_scale_k: 1.0,
        }
    }

    fn dsv4_module() -> Dsv4ModuleOp {
        Dsv4ModuleOp {
            name: "dsv4_module".into(),
            scale_factor: 1.0,
            attn_kind: AttnKind::Hca,
            num_heads: 128,
            native_heads: 128,
            tp_size: 1,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
            fmha_quant_mode: FmhaQuantMode::Fp8,
            gemm_quant_mode: GemmQuantMode::Fp8Block,
            architecture: "DeepseekV4ForCausalLM".into(),
            cp_size: 1,
            window_size: None,
            hidden_size: 7168,
            q_lora_rank: 1536,
            o_lora_rank: 1024,
            head_dim: 512,
            rope_head_dim: 64,
            index_n_heads: 64,
            index_head_dim: 128,
            index_topk: 1024,
            o_groups: Some(16),
        }
    }

    fn dsv4_megamoe() -> Dsv4MegaMoeOp {
        Dsv4MegaMoeOp {
            name: "context_megamoe".into(),
            scale_factor: 61.0,
            hidden_size: 7168,
            inter_size: 3072,
            topk: 6,
            num_experts: 384,
            moe_tp_size: 1,
            moe_ep_size: 8,
            quant_mode: MoeQuantMode::W4a8Mxfp4Mxfp8,
            workload_distribution: "balanced".into(),
            is_context: true,
            source_policy: "random".into(),
            pre_dispatch: "sglang_jit".into(),
            num_fused_shared_experts: 0,
            kernel_source: "deepgemm_megamoe".into(),
            kernel_dtype: "fp8_fp4".into(),
        }
    }

    fn mhc() -> MhcModuleOp {
        MhcModuleOp {
            name: "mhc_module".into(),
            scale_factor: 1.0,
            op: "pre".into(),
            hc_mult: 4,
            hidden_size: 7168,
            architecture: "DeepseekV4ForCausalLM".into(),
            sinkhorn_iters: 20,
            quant_mode: GemmQuantMode::Bfloat16,
            seq_split: 1,
        }
    }

    fn mamba2() -> Mamba2Op {
        Mamba2Op {
            name: "mamba2".into(),
            scale_factor: 1.0,
            kernel_source: "mamba_chunk_scan".into(),
            phase: "context".into(),
            d_model: 4096,
            d_state: 128,
            d_conv: 4,
            nheads: 128,
            head_dim: 64,
            n_groups: 8,
            chunk_size: 256,
        }
    }

    fn gdn() -> GdnOp {
        GdnOp {
            name: "gdn".into(),
            scale_factor: 1.0,
            kernel_source: "gdn_kernel".into(),
            phase: "generation".into(),
            d_model: 4096,
            d_conv: 4,
            num_k_heads: 16,
            head_k_dim: 128,
            num_v_heads: 32,
            head_v_dim: 128,
            // Non-default value so the round-trip notices a
            // `#[serde(default)]` swallowing the carried field.
            mamba_ssm_dtype: "bfloat16".into(),
        }
    }

    fn kda() -> KdaOp {
        KdaOp {
            name: "kda".into(),
            scale_factor: 1.0,
            kernel_source: "fused_sigmoid_gating_delta_rule_update".into(),
            phase: "verify".into(),
            d_model: 7168,
            d_conv: 4,
            num_k_heads: 16,
            head_k_dim: 128,
            num_v_heads: 16,
            head_v_dim: 128,
            draft_tokens: 4,
        }
    }

    fn wideep_context_mla() -> WideEpContextMlaOp {
        WideEpContextMlaOp {
            name: "wideep_context_mla".into(),
            scale_factor: 1.0,
            num_heads: 128,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
            fmha_quant_mode: FmhaQuantMode::Fp8,
            attn_backend: "flashinfer".into(),
            cp_size: 1,
        }
    }

    fn wideep_generation_mla() -> WideEpGenerationMlaOp {
        WideEpGenerationMlaOp {
            name: "wideep_generation_mla".into(),
            scale_factor: 1.0,
            num_heads: 128,
            kv_cache_dtype: KvCacheQuantMode::Fp8,
            fmha_quant_mode: FmhaQuantMode::Fp8,
            attn_backend: "flashinfer".into(),
        }
    }

    /// Large-EP comm phase with every optional field set to a NON-default
    /// value, so the round-trip would notice a `#[serde(default)]` swallowing
    /// a carried field.
    fn moe_all_to_all() -> MoeAllToAllOp {
        MoeAllToAllOp {
            name: "moe_dispatch".into(),
            scale_factor: 61.0,
            phase: "dispatch".into(),
            comm_backend: "deepep_ht".into(),
            comm_dtype: "fp8_block".into(),
            hidden_size: 7168,
            topk: 8,
            num_experts: 256,
            moe_ep_size: 16,
            node_num: 2,
            sms: 24,
            attention_tp_size: 2,
        }
    }

    /// Large-EP expert compute. `num_slots` / `kernel_source` are `Some(...)`
    /// here on purpose — the `None` (Python-default) case is what production
    /// emits, and both encodings must survive the wire.
    fn moe_expert_compute() -> MoeExpertComputeOp {
        MoeExpertComputeOp {
            name: "moe".into(),
            scale_factor: 61.0,
            hidden_size: 7168,
            inter_size: 2048,
            topk: 8,
            num_experts: 256,
            moe_ep_size: 16,
            quant_mode: MoeQuantMode::Fp8Block,
            workload_distribution: "power_law_1.2".into(),
            attention_dp_size: 8,
            inference_phase: "context".into(),
            num_slots: Some(288),
            kernel_source: Some("deepep_moe".into()),
            is_gated: true,
            enable_eplb: true,
        }
    }

    fn overlap() -> OverlapOp {
        // Recursive: nested children on both groups.
        OverlapOp {
            name: "overlap_attn_moe".into(),
            group_a: vec![OpSpec::ContextMla(context_mla()), OpSpec::Gemm(gemm())],
            group_b: vec![OpSpec::Moe(moe()), OpSpec::MoeDispatch(moe_dispatch())],
        }
    }

    fn fpm_forward() -> crate::operators::FpmForwardOp {
        // Recursive like Overlap/Fallback: sol_ops carries the model's
        // original granular list, so the round-trip must preserve nesting.
        crate::operators::FpmForwardOp {
            name: "fpm_forward_prefill".into(),
            phase: crate::operators::FpmPhase::Prefill,
            model_path: "org/model-a".into(),
            match_identity: vec![
                "nvfp4".into(),
                "nvfp4".into(),
                "bfloat16".into(),
                "half".into(),
                "fp8".into(),
                "4".into(),
                "1".into(),
                "1".into(),
                "4".into(),
                "1".into(),
                "1".into(),
            ],
            weight_bytes: 1.5e10,
            sol_ops: vec![
                OpSpec::Gemm(gemm()),
                OpSpec::ContextAttention(context_attention()),
            ],
        }
    }

    fn fallback() -> FallbackOp {
        // Recursive: a primary module op with a granular per-kernel fallback
        // chain that itself contains a nested Overlap.
        FallbackOp {
            name: "mla_fallback".into(),
            primary: Box::new(OpSpec::MlaModuleContext(mla_module())),
            fallback: vec![
                OpSpec::Gemm(gemm()),
                OpSpec::ContextMla(context_mla()),
                OpSpec::Overlap(overlap()),
            ],
        }
    }

    /// Every `Op` variant, constructed once. The exhaustive `match` below the
    /// `Vec` build forces the compiler to flag any newly added variant that
    /// this round-trip suite forgot to cover.
    fn all_op_variants() -> Vec<OpSpec> {
        let ops = vec![
            OpSpec::Gemm(gemm()),
            OpSpec::Embedding(embedding()),
            OpSpec::Elementwise(elementwise()),
            OpSpec::ContextAttention(context_attention()),
            OpSpec::GenerationAttention(generation_attention()),
            OpSpec::EncoderAttention(encoder_attention()),
            OpSpec::ContextMla(context_mla()),
            OpSpec::GenerationMla(generation_mla()),
            OpSpec::MlaModuleContext(mla_module()),
            OpSpec::MlaModuleGeneration(mla_module()),
            OpSpec::MlaBmm(mla_bmm()),
            OpSpec::Moe(moe()),
            OpSpec::MoeDispatch(moe_dispatch()),
            OpSpec::CustomAllReduce(custom_all_reduce()),
            OpSpec::Nccl(nccl()),
            OpSpec::P2P(p2p()),
            OpSpec::Vision(vision()),
            OpSpec::DsaContext(dsa_module()),
            OpSpec::DsaGeneration(dsa_module()),
            OpSpec::MsaContext(msa_module()),
            OpSpec::MsaGeneration(msa_module()),
            OpSpec::Dsv4Context(dsv4_module()),
            OpSpec::Dsv4Generation(dsv4_module()),
            OpSpec::Mhc(mhc()),
            OpSpec::Mamba2(mamba2()),
            OpSpec::Gdn(gdn()),
            OpSpec::WideEpContextMla(wideep_context_mla()),
            OpSpec::WideEpGenerationMla(wideep_generation_mla()),
            OpSpec::Overlap(overlap()),
            OpSpec::Fallback(fallback()),
            // Appended AFTER Fallback (bincode enum indices are positional;
            // appending shifts nothing, so no ENGINE_SPEC_SCHEMA_VERSION bump).
            OpSpec::Dsv4MegaMoe(dsv4_megamoe()),
            // Appended in wire order: Kda, FpmForward, then this PR's
            // large-EP pair.
            OpSpec::Kda(kda()),
            OpSpec::FpmForward(fpm_forward()),
            OpSpec::MoeAllToAll(moe_all_to_all()),
            OpSpec::MoeExpertCompute(moe_expert_compute()),
        ];

        // Exhaustiveness guard: if a variant is added to `Op`, this match
        // fails to compile until it is also added to `all_op_variants`.
        for op in &ops {
            match op {
                OpSpec::Gemm(_)
                | OpSpec::Embedding(_)
                | OpSpec::Elementwise(_)
                | OpSpec::ContextAttention(_)
                | OpSpec::GenerationAttention(_)
                | OpSpec::EncoderAttention(_)
                | OpSpec::ContextMla(_)
                | OpSpec::GenerationMla(_)
                | OpSpec::MlaModuleContext(_)
                | OpSpec::MlaModuleGeneration(_)
                | OpSpec::MlaBmm(_)
                | OpSpec::Moe(_)
                | OpSpec::MoeDispatch(_)
                | OpSpec::CustomAllReduce(_)
                | OpSpec::Nccl(_)
                | OpSpec::P2P(_)
                | OpSpec::Vision(_)
                | OpSpec::DsaContext(_)
                | OpSpec::DsaGeneration(_)
                | OpSpec::MsaContext(_)
                | OpSpec::MsaGeneration(_)
                | OpSpec::Dsv4Context(_)
                | OpSpec::Dsv4Generation(_)
                | OpSpec::Mhc(_)
                | OpSpec::Mamba2(_)
                | OpSpec::Gdn(_)
                | OpSpec::WideEpContextMla(_)
                | OpSpec::WideEpGenerationMla(_)
                | OpSpec::FpmForward(_)
                | OpSpec::Overlap(_)
                | OpSpec::Fallback(_)
                | OpSpec::Dsv4MegaMoe(_)
                | OpSpec::Kda(_)
                | OpSpec::MoeAllToAll(_)
                | OpSpec::MoeExpertCompute(_) => {}
            }
        }
        ops
    }

    fn sample_engine_config() -> EngineConfig {
        EngineConfig {
            schema_version: ENGINE_CONFIG_SCHEMA_VERSION,
            model_name: "deepseek-ai/DeepSeek-V3".into(),
            system_name: "h200_sxm".into(),
            systems_path: None,
            backend: crate::BackendKind::Trtllm,
            backend_version: Some("1.0.0rc3".into()),
            forward_model: None,
            kv_block_size: Some(64),
            parallel: ParallelMapping {
                tp_size: 8,
                pp_size: 1,
                attention_dp_size: Some(8),
                moe_tp_size: Some(1),
                moe_ep_size: Some(8),
                cp_size: None,
            },
            quantization: QuantizationConfig {
                weight_dtype: Some(DataType::Fp8),
                moe_dtype: Some(DataType::Fp8),
                activation_dtype: Some(DataType::Fp8),
                kv_cache_dtype: Some(DataType::Fp8),
            },
            speculative: Some(SpeculativeConfig { nextn: Some(1) }),
            enable_shared_layer: None,
            strict_provenance: false,
            database_mode: Default::default(),
            tolerate_dirless_version: false,
            transfer_policy: None,
            extra: BTreeMap::new(),
        }
    }

    /// Pin the bincode POSITIONAL variant index of the first and the two last
    /// `Op` variants. bincode encodes an enum as a leading 4-byte LE variant
    /// index, so inserting or removing a variant mid-enum silently reinterprets
    /// every later variant on the wire.
    ///
    /// These pin bincode positional indices; if this test fails you reordered/
    /// inserted mid-enum — append instead, or bump ENGINE_SPEC_SCHEMA_VERSION
    /// in lockstep (config.rs + engine.py).
    #[test]
    fn op_variant_indices_are_pinned() {
        const GEMM_INDEX: u32 = 0;
        // Re-derived after current main's Kda and FpmForward tail variants,
        // and after retiring the two mid-enum wideEP MoE variants.
        const MOE_ALL_TO_ALL_INDEX: u32 = 33;
        const MOE_EXPERT_COMPUTE_INDEX: u32 = 34;

        let index_of = |op: &OpSpec| -> u32 {
            let bytes = bincode::serialize(op).expect("serialize op");
            u32::from_le_bytes(bytes[..4].try_into().expect("4-byte variant index prefix"))
        };

        assert_eq!(
            index_of(&OpSpec::Gemm(gemm())),
            GEMM_INDEX,
            "first variant moved"
        );
        assert_eq!(
            index_of(&OpSpec::MoeAllToAll(moe_all_to_all())),
            MOE_ALL_TO_ALL_INDEX,
            "MoeAllToAll index moved"
        );
        assert_eq!(
            index_of(&OpSpec::MoeExpertCompute(moe_expert_compute())),
            MOE_EXPERT_COMPUTE_INDEX,
            "MoeExpertCompute index moved"
        );

        // The two last variants must stay adjacent and terminal: appending is
        // the only safe growth direction.
        assert_eq!(MOE_EXPERT_COMPUTE_INDEX, MOE_ALL_TO_ALL_INDEX + 1);
        assert_eq!(
            MOE_EXPERT_COMPUTE_INDEX as usize + 1,
            all_op_variants().len(),
            "all_op_variants() must cover exactly the pinned variant count"
        );
    }

    #[test]
    fn every_op_variant_round_trips_through_bincode() {
        for op in all_op_variants() {
            let bytes = bincode::serialize(&op).expect("serialize op");
            let decoded: OpSpec = bincode::deserialize(&bytes).expect("deserialize op");
            assert_eq!(op, decoded, "round-trip mismatch for {:?}", op);
        }
    }

    #[test]
    fn recursive_overlap_round_trips_with_nested_children() {
        let op = OpSpec::Overlap(overlap());
        let bytes = bincode::serialize(&op).unwrap();
        let decoded: OpSpec = bincode::deserialize(&bytes).unwrap();
        assert_eq!(op, decoded);
    }

    #[test]
    fn recursive_fallback_round_trips_with_nested_children() {
        let op = OpSpec::Fallback(fallback());
        let bytes = bincode::serialize(&op).unwrap();
        let decoded: OpSpec = bincode::deserialize(&bytes).unwrap();
        assert_eq!(op, decoded);
    }

    #[test]
    fn mla_module_none_native_round_trips_followed_by_another_op() {
        // native_num_heads=None must still be serialized (no skip_serializing_if):
        // bincode decodes positionally, so an omitted Option would desync the
        // ops decoded after it (#1458 review).
        let mut none_native = mla_module();
        none_native.native_num_heads = None;
        let spec = EngineSpec::new(
            sample_engine_config(),
            vec![OpSpec::MlaModuleContext(none_native), OpSpec::Gemm(gemm())],
            vec![
                OpSpec::MlaModuleGeneration(mla_module()),
                OpSpec::Moe(moe()),
            ],
        );
        let bytes = spec.to_bincode().expect("to_bincode");
        let decoded = EngineSpec::from_bincode(&bytes).expect("from_bincode");
        assert_eq!(spec, decoded);
    }

    #[test]
    fn engine_spec_round_trips_through_bincode() {
        let spec = EngineSpec::new(
            sample_engine_config(),
            vec![
                OpSpec::Embedding(embedding()),
                OpSpec::ContextAttention(context_attention()),
                OpSpec::Overlap(overlap()),
                OpSpec::Gemm(gemm()),
            ],
            vec![
                OpSpec::GenerationAttention(generation_attention()),
                OpSpec::Fallback(fallback()),
                OpSpec::Moe(moe()),
            ],
        );

        assert_eq!(spec.schema_version, ENGINE_SPEC_SCHEMA_VERSION);

        let bytes = spec.to_bincode().expect("to_bincode");
        let decoded = EngineSpec::from_bincode(&bytes).expect("from_bincode");
        assert_eq!(spec, decoded);
    }

    /// A version skew combined with an op-layout change must surface as a clear
    /// [`AicError::UnsupportedSchemaVersion`], NOT a generic bincode I/O error.
    ///
    /// This reproduces the cross-version failure mode: a producer at a different
    /// schema version emits op payloads whose layout the consumer cannot decode.
    /// We simulate it by stamping a foreign version into the leading prefix and
    /// truncating the op payload. `from_bincode` must read + reject the version
    /// *before* it attempts to decode the (now-undecodable) op lists.
    #[test]
    fn version_skew_reports_unsupported_before_payload_decode() {
        let spec = EngineSpec::new(
            sample_engine_config(),
            vec![
                OpSpec::Gemm(gemm()),
                OpSpec::ContextAttention(context_attention()),
            ],
            vec![OpSpec::GenerationAttention(generation_attention())],
        );
        let mut bytes = spec.to_bincode().expect("to_bincode");

        // Overwrite the 4-byte little-endian `schema_version` prefix with a
        // version this consumer does not speak.
        let foreign = ENGINE_SPEC_SCHEMA_VERSION + 1;
        bytes[..4].copy_from_slice(&foreign.to_le_bytes());
        // Corrupt the op payload so a decode-first implementation fails there
        // with a generic bincode I/O error instead of reaching the version gate.
        bytes.truncate(bytes.len() - 8);

        match EngineSpec::from_bincode(&bytes) {
            Err(AicError::UnsupportedSchemaVersion {
                kind,
                got,
                expected,
            }) => {
                assert_eq!(kind, "EngineSpec");
                assert_eq!(got, foreign);
                assert_eq!(expected, ENGINE_SPEC_SCHEMA_VERSION);
            }
            other => {
                panic!("expected UnsupportedSchemaVersion before payload decode, got {other:?}")
            }
        }
    }

    /// Canonical valid spec used by the handshake tests below.
    fn handshake_spec() -> EngineSpec {
        EngineSpec::new(
            sample_engine_config(),
            vec![
                OpSpec::Gemm(gemm()),
                OpSpec::ContextAttention(context_attention()),
            ],
            vec![OpSpec::GenerationAttention(generation_attention())],
        )
    }

    /// Round-trip preserves the stamped schema version end to end.
    #[test]
    fn from_bincode_round_trips_and_preserves_version() {
        let spec = handshake_spec();
        let decoded = EngineSpec::from_bincode(&spec.to_bincode().expect("to_bincode"))
            .expect("from_bincode");
        assert_eq!(decoded, spec);
        assert_eq!(decoded.schema_version, ENGINE_SPEC_SCHEMA_VERSION);
    }

    /// A buffer too short to even hold the 4-byte version prefix must fail at the
    /// prefix stage, not deep in the (absent) payload.
    #[test]
    fn from_bincode_rejects_empty_buffer() {
        match EngineSpec::from_bincode(&[]) {
            Err(AicError::EngineSpec(msg)) => {
                assert!(
                    msg.contains("schema_version prefix"),
                    "message should name the prefix stage, got: {msg}"
                );
            }
            other => panic!("expected EngineSpec prefix error, got {other:?}"),
        }
    }

    /// The version gate fires even when the op payload is fully intact — the
    /// rejection is driven by the version alone, not by a decode failure.
    #[test]
    fn from_bincode_version_gate_fires_with_intact_payload() {
        let mut bytes = handshake_spec().to_bincode().expect("to_bincode");
        let foreign = ENGINE_SPEC_SCHEMA_VERSION + 7;
        bytes[..4].copy_from_slice(&foreign.to_le_bytes()); // only the prefix changes

        match EngineSpec::from_bincode(&bytes) {
            Err(AicError::UnsupportedSchemaVersion {
                kind,
                got,
                expected,
            }) => {
                assert_eq!(kind, "EngineSpec");
                assert_eq!(got, foreign);
                assert_eq!(expected, ENGINE_SPEC_SCHEMA_VERSION);
            }
            other => panic!("expected UnsupportedSchemaVersion, got {other:?}"),
        }
    }

    /// v11 -> v12 regression (PR-6): `DsaModuleOp` gained
    /// `attn_projection_quant_modes`, a positional bincode layout change. A
    /// pre-PR v11 producer's DSA payload must be rejected by the VERSION GATE
    /// (before any op decoding) as `UnsupportedSchemaVersion` — never reach
    /// the op-payload stage where the missing Option tag would surface as an
    /// opaque "unexpected end of file".
    #[test]
    fn from_bincode_rejects_v11_dsa_producer_at_the_version_gate() {
        let spec = EngineSpec::new(
            sample_engine_config(),
            vec![OpSpec::DsaContext(dsa_module())],
            vec![],
        );
        let mut bytes = spec.to_bincode().expect("to_bincode");
        bytes[..4].copy_from_slice(&11u32.to_le_bytes()); // a v11 producer's stamp

        match EngineSpec::from_bincode(&bytes) {
            Err(AicError::UnsupportedSchemaVersion {
                kind,
                got,
                expected,
            }) => {
                assert_eq!(kind, "EngineSpec");
                assert_eq!(got, 11);
                assert_eq!(expected, ENGINE_SPEC_SCHEMA_VERSION);
            }
            other => {
                panic!("expected UnsupportedSchemaVersion for a v11 DSA payload, got {other:?}")
            }
        }
    }

    /// v13 -> v14 regression (PR #1533): `GdnOp` gained
    /// `mamba_ssm_dtype`, a positional bincode layout change. A pre-PR v13
    /// producer's GDN payload must be rejected by the version gate before the
    /// missing trailing string reaches op decoding as an opaque EOF.
    #[test]
    fn from_bincode_rejects_v13_gdn_producer_at_the_version_gate() {
        let mut legacy_gdn = gdn();
        legacy_gdn.mamba_ssm_dtype.clear();
        let spec = EngineSpec::new(
            sample_engine_config(),
            vec![],
            vec![OpSpec::Gdn(legacy_gdn)],
        );
        let mut bytes = spec.to_bincode().expect("to_bincode");

        // The GDN op is the final value in the wire payload. An empty String
        // is encoded as its 8-byte length, so removing that suffix recreates
        // the exact pre-field GdnOp layout emitted by a v13 producer.
        assert_eq!(&bytes[bytes.len() - 8..], &[0; 8]);
        bytes.truncate(bytes.len() - 8);
        bytes[..4].copy_from_slice(&13u32.to_le_bytes());

        match EngineSpec::from_bincode(&bytes) {
            Err(AicError::UnsupportedSchemaVersion {
                kind,
                got,
                expected,
            }) => {
                assert_eq!(kind, "EngineSpec");
                assert_eq!(got, 13);
                assert_eq!(expected, ENGINE_SPEC_SCHEMA_VERSION);
            }
            other => {
                panic!("expected UnsupportedSchemaVersion for a v13 GDN payload, got {other:?}")
            }
        }
    }

    /// A correct version but an undecodable op payload is NOT a version skew: it
    /// must surface as an op-payload-stage `EngineSpec` error (op-layout drift
    /// within a version, or corruption), naming the stage and the version.
    #[test]
    fn from_bincode_matching_version_corrupt_ops_names_op_payload_stage() {
        let mut bytes = handshake_spec().to_bincode().expect("to_bincode");
        // Leave the version prefix intact; corrupt the trailing op payload.
        bytes.truncate(bytes.len() - 8);

        match EngineSpec::from_bincode(&bytes) {
            Err(AicError::EngineSpec(msg)) => {
                assert!(
                    msg.contains("op payloads"),
                    "message should name the op-payload stage, got: {msg}"
                );
                assert!(
                    msg.contains(&ENGINE_SPEC_SCHEMA_VERSION.to_string()),
                    "message should cite the matching version, got: {msg}"
                );
            }
            other => panic!("expected EngineSpec op-payload error, got {other:?}"),
        }
    }

    /// A well-formed wire buffer whose embedded `engine_json` is not valid JSON
    /// must fail at the engine-JSON stage (after the version gate and op decode
    /// both pass), naming that stage.
    #[test]
    fn from_bincode_invalid_engine_json_names_json_stage() {
        // Hand-build a wire buffer with the current version, empty op lists, and
        // a deliberately malformed `engine_json`.
        let wire = BincodeWire {
            schema_version: ENGINE_SPEC_SCHEMA_VERSION,
            engine_json: "this is not json".to_string(),
            context_ops: vec![],
            generation_ops: vec![],
        };
        let bytes = bincode::serialize(&wire).expect("serialize wire");

        match EngineSpec::from_bincode(&bytes) {
            Err(AicError::EngineSpec(msg)) => {
                assert!(
                    msg.contains("engine JSON"),
                    "message should name the engine-JSON stage, got: {msg}"
                );
            }
            other => panic!("expected EngineSpec engine-JSON error, got {other:?}"),
        }
    }
}