vyre-driver-cuda 0.7.0

use std::ffi::c_void;
use std::sync::Arc;

use smallvec::SmallVec;
use vyre_driver::binding::BindingRole;
use vyre_driver::{BackendError, DispatchConfig, PendingDispatch};
use vyre_foundation::ir::Program;

use crate::backend::allocations::{DispatchAllocations, HostTransferAllocations};
use crate::backend::copy::aligned_async_copy_len;
use crate::backend::dispatch::CudaBackend;
use crate::backend::dispatch_phase_probe as probe;
use crate::backend::launch_params::launch_param_byte_len;
use crate::backend::module_cache::ModuleCacheKey;
use crate::backend::ordering::sort_unstable_by_key_if_needed;
use crate::backend::output_range::{cuda_output_readback_for_binding, CudaOutputReadback};
use crate::backend::plan::CudaDispatchPlan;
use crate::backend::resident::{
    resident_bindings_from_handles, CudaDispatchBinding, CudaResidentBuffer, ResidentViewCache,
};
use crate::backend::resident_dispatch::helpers::{
    enqueue_optional_resident_h2d_copy, enqueue_resident_h2d_copy, next_dispatch_binding,
    resident_required_handles, validate_dense_resident_output_indices,
};
use crate::backend::resident_dispatch_support::{
    add_resident_dispatch_bytes, add_resident_dispatch_u64_count, CudaResidentDispatch,
};
use crate::backend::staging_reserve::{reserve_smallvec, reserved_vec};
use crate::numeric::CUDA_NUMERIC;

pub(super) fn resident_output_clear_for_readback(
    base_ptr: u64,
    readback: CudaOutputReadback,
    binding_name: &str,
) -> Result<Option<(u64, usize)>, BackendError> {
    if readback.byte_len == 0 {
        return Ok(None);
    }
    let clear_ptr = vyre_driver::accounting::checked_add_u64_usize_offset_lazy(
        base_ptr,
        readback.device_offset,
        || {
            BackendError::InvalidProgram {
            fix: format!(
                "Fix: CUDA resident output clear offset {} for binding `{binding_name}` does not fit CUdeviceptr arithmetic.",
                readback.device_offset
            ),
        }
        },
        || {
            BackendError::InvalidProgram {
            fix: format!(
                "Fix: CUDA resident output clear pointer for binding `{binding_name}` overflowed at offset {}.",
                readback.device_offset
            ),
        }
        },
    )?;
    Ok(Some((clear_ptr, readback.byte_len)))
}

impl CudaBackend {
    /// Dispatch a Program asynchronously using caller-provided CUDA-resident buffers.
    pub fn dispatch_resident_async(
        &self,
        program: &Program,
        handles: &[CudaResidentBuffer],
        config: &DispatchConfig,
    ) -> Result<Box<dyn PendingDispatch>, BackendError> {
        self.dispatch_bindings_async(program, &resident_bindings_from_handles(handles)?, config)
    }

    /// Dispatch a Program asynchronously against a mixed binding list.
    ///
    /// Residency is per binding: resident entries bind their existing device
    /// memory, borrowed entries are staged into the transient pool for this
    /// dispatch alone.
    pub(crate) fn dispatch_bindings_async(
        &self,
        program: &Program,
        bindings: &[CudaDispatchBinding<'_>],
        config: &DispatchConfig,
    ) -> Result<Box<dyn PendingDispatch>, BackendError> {
        if crate::instrumentation::cuda_resident_borrowed_fallback_enabled() {
            let outputs = self.dispatch_resident_via_borrowed(program, bindings, config)?;
            return Ok(Box::new(crate::stream::CudaPendingDispatch::new_ready(
                Arc::clone(&self.ctx),
                Arc::clone(&self.launch_resources),
                outputs,
                Arc::clone(&self.telemetry),
            )));
        }
        let prepared = self.prepare_resident_dispatch(program, bindings, config)?;
        let (ptx_src, ptx_source_key) = self.ptx_for_program_cached_with_key(program, config)?;
        let module_key = self.module_cache_key_for_ptx_source_key(ptx_source_key)?;
        let native = self.dispatch_resident_async_concrete_with_ptx_key(
            program, bindings, config, &ptx_src, module_key, false, None, true, &prepared,
        )?;
        Ok(Box::new(native.pending))
    }

    #[allow(clippy::too_many_arguments)]
    pub(crate) fn dispatch_resident_async_concrete_with_ptx_key(
        &self,
        program: &Program,
        bindings: &[CudaDispatchBinding<'_>],
        _config: &DispatchConfig,
        ptx_src: &str,
        module_key: ModuleCacheKey,
        capture_timing: bool,
        static_params_ptr: Option<u64>,
        capture_outputs: bool,
        prepared: &CudaDispatchPlan,
    ) -> Result<CudaResidentDispatch, BackendError> {
        let _profiler_range =
            crate::profiler::cuda_profiler_range(crate::profiler::CUDA_RESIDENT_DISPATCH_RANGE);
        let trace = crate::instrumentation::cuda_stage_trace_enabled();
        let start = std::time::Instant::now();
        if trace {
            tracing::debug!(
                "[cuda-trace] resident dispatch start buffers={} bindings={}",
                program.buffers().len(),
                bindings.len()
            );
        }
        self.warmup()?;
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident warmup",
                start.elapsed().as_millis()
            );
        }
        let required_bindings = resident_required_handles(prepared)?;
        if bindings.len() != required_bindings {
            return Err(BackendError::InvalidProgram {
                fix: format!(
                    "Fix: CUDA resident dispatch expected {required_bindings} bound resource(s) but received {}.",
                    bindings.len()
                ),
            });
        }
        // Borrowed bindings are staged out of the transient pool, so they have
        // to clear the same live-device budget a fully borrowed dispatch does.
        // An all-resident dispatch stages nothing and skips the preflight
        // entirely, so the existing hot path gains no driver calls.
        self.validate_mixed_dispatch_staging_budget(
            prepared,
            bindings,
            "CUDA mixed resident dispatch",
        )?;
        let mut allocations =
            DispatchAllocations::new(program.buffers().len(), Arc::clone(&self.transient_pool))?;
        let mut launch_ptrs = SmallVec::<[u64; 8]>::new();
        reserve_smallvec(
            &mut launch_ptrs,
            prepared.bindings.bindings.len(),
            "resident dispatch launch pointers",
        )?;
        let mut output_stage_readbacks = SmallVec::<[(u64, CudaOutputReadback); 8]>::new();
        reserve_smallvec(
            &mut output_stage_readbacks,
            if capture_outputs {
                prepared.output_binding_indices.len()
            } else {
                0
            },
            "resident dispatch output staged readbacks",
        )?;
        let mut next_binding = 0usize;
        let mut output_handles_by_index =
            SmallVec::<[(usize, Option<CudaResidentBuffer>, CudaOutputReadback, u64); 8]>::new();
        reserve_smallvec(
            &mut output_handles_by_index,
            prepared.output_binding_indices.len(),
            "resident dispatch output handles by index",
        )?;
        let mut output_clears = SmallVec::<[(u64, usize); 8]>::new();
        reserve_smallvec(
            &mut output_clears,
            prepared.output_binding_indices.len(),
            "resident dispatch output clears",
        )?;
        let mut resident_view_cache = ResidentViewCache::new();
        reserve_smallvec(
            &mut resident_view_cache,
            bindings.len(),
            "resident dispatch view cache",
        )?;
        let mut resident_handles = SmallVec::<[CudaResidentBuffer; 8]>::new();
        reserve_smallvec(
            &mut resident_handles,
            bindings.len(),
            "resident dispatch in-flight handles",
        )?;
        let mut borrowed_stages = SmallVec::<[(u64, &[u8]); 8]>::new();
        reserve_smallvec(
            &mut borrowed_stages,
            bindings.len(),
            "resident dispatch borrowed staging",
        )?;
        for binding in &prepared.bindings.bindings {
            if binding.role == BindingRole::Shared {
                continue;
            }
            let source =
                next_dispatch_binding(bindings, &mut next_binding, "resident dispatch launch")?;
            let (launch_ptr, bound_byte_len, resident_handle) = match source {
                CudaDispatchBinding::Resident(handle) => {
                    let resident = self.resident_store.view_cached(
                        handle,
                        &mut resident_view_cache,
                        "resident dispatch view cache",
                    )?;
                    if let Some(expected) = binding.static_byte_len {
                        if resident.byte_len < expected {
                            return Err(BackendError::InvalidProgram {
                                fix: format!(
                                    "Fix: CUDA resident buffer `{}` expected at least {expected} bytes but handle {} has {} bytes.",
                                    binding.name, handle.handle, resident.byte_len
                                ),
                            });
                        }
                    }
                    if resident.ptr == 0 {
                        return Err(BackendError::InvalidProgram {
                            fix: format!(
                                "Fix: CUDA resident binding `{}` resolved to a null device pointer; resident launch arguments must preserve descriptor order.",
                                binding.name
                            ),
                        });
                    }
                    resident_handles.push(handle);
                    (resident.ptr, resident.byte_len, Some(handle))
                }
                CudaDispatchBinding::Borrowed(bytes) => {
                    // An output-only binding carries no host payload to stage,
                    // so it is sized from the declared output extent exactly
                    // like the fully borrowed dispatch sizes it.
                    let staged_byte_len = match binding.input_index {
                        Some(_) => bytes.len(),
                        None => binding.static_byte_len.ok_or_else(|| {
                            BackendError::InvalidProgram {
                            fix: format!(
                                "Fix: CUDA borrowed output `{}` needs a static byte length before launch; set BufferDecl::with_count or output_byte_range, or bind a resident buffer for that output.",
                                binding.name
                            ),
                        }
                        })?,
                    };
                    if let Some(expected) = binding.static_byte_len {
                        if staged_byte_len < expected {
                            return Err(BackendError::InvalidProgram {
                                fix: format!(
                                    "Fix: CUDA borrowed binding `{}` expected at least {expected} bytes but the supplied buffer has {staged_byte_len} bytes.",
                                    binding.name
                                ),
                            });
                        }
                    }
                    let allocation = self
                        .transient_pool
                        .acquire(aligned_async_copy_len(staged_byte_len)?)?;
                    self.telemetry
                        .record_transient_allocation_bytes(CUDA_NUMERIC.usize_to_u64(
                            allocation.byte_len,
                            "resident dispatch borrowed staging byte count",
                        )?);
                    let staged_ptr = allocation.ptr;
                    allocations.set_ptr(binding.buffer_index, allocation, &binding.name)?;
                    if binding.input_index.is_some() && !bytes.is_empty() {
                        borrowed_stages.push((staged_ptr, bytes));
                    }
                    (staged_ptr, staged_byte_len, None)
                }
            };
            launch_ptrs.push(launch_ptr);
            if let Some(output_index) = binding.output_index {
                let full_byte_len = match binding.static_byte_len {
                    Some(len) => len,
                    None => bound_byte_len,
                };
                let readback = cuda_output_readback_for_binding(
                    program.buffers(),
                    binding.buffer_index,
                    &binding.name,
                    full_byte_len,
                    "resident async output readback",
                )?;
                output_handles_by_index.push((output_index, resident_handle, readback, launch_ptr));
                if binding.input_index.is_none() {
                    output_clears.extend(resident_output_clear_for_readback(
                        launch_ptr,
                        readback,
                        &binding.name,
                    )?);
                }
            }
        }
        if output_handles_by_index.len() != prepared.output_binding_indices.len() {
            return Err(BackendError::InvalidProgram {
                fix: format!(
                    "Fix: CUDA resident dispatch expected {} output handle(s) but resolved {}.",
                    prepared.output_binding_indices.len(),
                    output_handles_by_index.len()
                ),
            });
        }
        sort_unstable_by_key_if_needed(
            output_handles_by_index.as_mut_slice(),
            |(output_index, _, _, _)| *output_index,
        );
        validate_dense_resident_output_indices(
            output_handles_by_index
                .iter()
                .map(|(output_index, _, _, _)| *output_index),
            prepared.output_binding_indices.len(),
            "resident dispatch output handles",
        )?;
        let mut output_handles = SmallVec::<[Option<CudaResidentBuffer>; 8]>::new();
        reserve_smallvec(
            &mut output_handles,
            output_handles_by_index.len(),
            "resident dispatch output handles",
        )?;
        let mut output_readbacks = SmallVec::<[CudaOutputReadback; 8]>::new();
        reserve_smallvec(
            &mut output_readbacks,
            output_handles_by_index.len(),
            "resident dispatch output readbacks",
        )?;
        for (_, handle, readback, launch_ptr) in output_handles_by_index {
            output_handles.push(handle);
            output_readbacks.push(readback);
            if capture_outputs {
                output_stage_readbacks.push((launch_ptr, readback));
            }
        }
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident args/readbacks launch_ptrs={:x?} output_clears={} output_stage_readbacks={}",
                start.elapsed().as_millis(),
                launch_ptrs,
                output_clears.len(),
                output_stage_readbacks.len()
            );
        }

        let param_bytes = launch_param_byte_len(&prepared.launch.param_words, "resident dispatch")?;
        let param_transfer_slots = usize::from(static_params_ptr.is_none() && param_bytes != 0);
        let transfer_capacity = borrowed_stages
            .len()
            .checked_add(param_transfer_slots)
            .ok_or_else(|| BackendError::InvalidProgram {
                fix: "Fix: CUDA resident dispatch host staging slot count overflowed usize; shard the dispatch before launch.".to_string(),
            })?;
        let mut host_transfers = HostTransferAllocations::with_capacity(
            Arc::clone(&self.host_pool),
            transfer_capacity,
            output_stage_readbacks.len(),
        )?;
        let mut borrowed_upload_copies = SmallVec::<[(u64, *const c_void, usize); 8]>::new();
        reserve_smallvec(
            &mut borrowed_upload_copies,
            borrowed_stages.len(),
            "resident dispatch borrowed upload copies",
        )?;
        let mut borrowed_upload_bytes = 0_u64;
        let mut borrowed_upload_ops = 0_u64;
        for &(staged_ptr, bytes) in &borrowed_stages {
            let copy_byte_len = aligned_async_copy_len(bytes.len())?;
            let host_ptr = host_transfers.push_upload_padded(bytes, copy_byte_len)?;
            add_resident_dispatch_bytes(
                &mut borrowed_upload_bytes,
                bytes.len(),
                "resident dispatch borrowed upload",
            )?;
            add_resident_dispatch_u64_count(
                &mut borrowed_upload_ops,
                "resident dispatch borrowed upload operation",
            )?;
            borrowed_upload_copies.push((staged_ptr, host_ptr, copy_byte_len));
        }
        let mut param_upload: Option<(u64, *const c_void, usize)> = None;
        let params_ptr = match static_params_ptr {
            Some(ptr) => ptr,
            None if param_bytes == 0 => 0,
            None => {
                let (params_ptr, upload) = self.prepare_resident_param_upload(
                    &prepared.launch.param_words,
                    param_bytes,
                    "CUDA resident dispatch parameter bytes",
                    "CUDA resident dispatch parameter upload",
                    "resident dispatch parameter allocation byte count",
                    "resident dispatch parameter upload byte count",
                    &mut allocations,
                    &mut host_transfers,
                )?;
                param_upload = upload;
                params_ptr
            }
        };
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident params ptr=0x{params_ptr:x} words={:?} grid={:?} workgroup={:?} element_count={}",
                start.elapsed().as_millis(),
                prepared.launch.param_words,
                prepared.launch.grid,
                prepared.launch.workgroup,
                prepared.launch.element_count
            );
        }

        let resident_use = self.resident_store.mark_inflight(&resident_handles)?;
        let launch_resources = crate::stream::CudaLaunchResourceLease::acquire(
            Arc::clone(&self.launch_resources),
            capture_timing,
        )?;
        let mut launch_resources = Some(launch_resources);
        let mut allocations = Some(allocations);
        let mut resident_use = Some(resident_use);
        let mut host_transfers = Some(host_transfers);
        let stream_raw = launch_resources
            .as_ref()
            .ok_or_else(|| BackendError::InvalidProgram {
                fix: "Fix: CUDA resident dispatch launch resources were consumed before enqueue; rebuild launch resource ownership before launching.".to_string(),
            })?
            .stream_raw()?;
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident allocations/stream",
                start.elapsed().as_millis()
            );
        }
        // Kernel-only device window, probe path only.
        //
        // The outer timing events below bracket host work as well: function
        // resolution, the grid-barrier lease (which scans the PTX text), and
        // the post-launch release (which synchronizes the stream and reads the
        // arrival counter back). CUDA event elapsed time is a difference of two
        // DEVICE timestamps, so every stretch where the stream drained while
        // the host had not yet enqueued the next item lands inside it. This
        // inner pair brackets the launch loop alone, which is what lets the
        // probe price that error instead of leaving it as a caveat on every
        // device figure derived from the outer pair.
        let kernel_events = if probe::enabled() && capture_timing {
            Some(self.launch_resources.acquire_timing_event_pair()?)
        } else {
            None
        };
        let enqueue_result = (|| {
            enqueue_optional_resident_h2d_copy(param_upload, stream_raw)?;
            for &(dst_ptr, host_ptr, byte_len) in &borrowed_upload_copies {
                enqueue_resident_h2d_copy(dst_ptr, host_ptr, byte_len, stream_raw)?;
            }
            if borrowed_upload_ops != 0 {
                self.telemetry
                    .record_host_to_device_bytes(borrowed_upload_bytes);
                self.telemetry
                    .record_host_upload_operations(borrowed_upload_ops);
            }
            if trace {
                tracing::debug!(
                    "[cuda-trace] +{}ms resident param upload enqueued",
                    start.elapsed().as_millis()
                );
            }
            for &(dst_ptr, byte_len) in &output_clears {
                // SAFETY: FFI to libcuda.so. Resident output pointers were
                // validated above and byte lengths come from the binding/readback
                // plan. The memset is enqueued on the same stream before launch,
                // matching the borrowed CUDA dispatch output-zeroing contract.
                unsafe {
                    crate::backend::copy::memset_d8_async_checked(
                        dst_ptr, 0, byte_len, stream_raw,
                    )?;
                }
            }
            if trace {
                tracing::debug!(
                    "[cuda-trace] +{}ms resident output clears enqueued",
                    start.elapsed().as_millis()
                );
            }
            if crate::instrumentation::cuda_resident_sync_before_launch_enabled() {
                // SAFETY: stream_raw is owned by launch_resources for the
                // duration of this dispatch. This opt-in diagnostic fence isolates
                // setup copies/memsets from kernel execution without changing the
                // release default.
                crate::stream::synchronize_raw_stream(
                    stream_raw,
                    "cuStreamSynchronize (resident prelaunch)",
                )?;
                self.telemetry.record_sync_point();
                if trace {
                    tracing::debug!(
                        "[cuda-trace] +{}ms resident prelaunch sync complete",
                        start.elapsed().as_millis()
                    );
                }
            }

            probe::charge_since(probe::Phase::Stage, start);
            if let Some((start_event, _)) = launch_resources
                .as_ref()
                .ok_or_else(|| BackendError::InvalidProgram {
                    fix: "Fix: CUDA resident dispatch launch resources were consumed before timing-event record.".to_string(),
                })?
                .timing_events()?
            {
                start_event.record(stream_raw)?;
            }
            // Fixpoint loop  -  see dispatch_borrowed_async_with_ptx_concrete
            // for the contract. Resolve the CUDA function and argument vector
            // once; fixpoint iterations are kernel replays, not relowering or
            // module-cache lookups.
            let resolve_started = probe::mark();
            let func = self.resolve_launch_function(
                ptx_src,
                module_key,
                &prepared.launch,
                prepared.cooperative,
            )?;
            if trace {
                tracing::debug!(
                    "[cuda-trace] +{}ms resident resolve_launch_function",
                    start.elapsed().as_millis()
                );
            }
            let mut params_ref = params_ptr;
            let mut kernel_args = Self::kernel_args(&mut launch_ptrs, &mut params_ref)?;
            probe::charge(probe::Phase::Resolve, resolve_started);
            // Hold this module's grid-barrier counter across the launch sequence.
            // The lease blocks while another cooperative launch of the same module
            // is still in flight; see `GridBarrierGate`.
            let lease_started = probe::mark();
            let grid_barrier = self.lease_grid_barrier(program, prepared, ptx_src, module_key)?;
            probe::charge(probe::Phase::Lease, lease_started);
            // `launch_then_release` runs the launches and ends the lease in the
            // one safe order: the release synchronizes the stream before freeing
            // the gate, so a launch failure cannot leave a grid spinning while the
            // next sequence resets the counter underneath it.
            let launch_and_release_started = probe::mark();
            grid_barrier.launch_then_release(
                stream_raw,
                "resident async dispatch grid-sync launch",
                |grid_barrier| {
                    if let Some((kernel_start, _)) = kernel_events.as_ref() {
                        kernel_start.record(stream_raw)?;
                    }
                    probe::measure(probe::Phase::LaunchLoop, || {
                        for _ in 0..prepared.fixpoint_iterations {
                            // SAFETY: stream_raw is the live resident dispatch
                            // stream and outlives the memset enqueued ahead of
                            // the launch.
                            unsafe {
                                grid_barrier.enqueue_reset(stream_raw)?;
                            }
                            self.launch_prevalidated_function(
                                func,
                                &mut kernel_args,
                                &prepared.launch,
                                stream_raw,
                                false,
                                prepared.cooperative,
                            )?;
                        }
                        Ok::<(), BackendError>(())
                    })?;
                    if let Some((_, kernel_end)) = kernel_events.as_ref() {
                        kernel_end.record(stream_raw)?;
                    }
                    Ok(())
                },
            )?;
            probe::charge_remainder(
                probe::Phase::Release,
                launch_and_release_started,
                probe::Phase::LaunchLoop,
            );
            if let Some((_, end_event)) = launch_resources
                .as_ref()
                .ok_or_else(|| BackendError::InvalidProgram {
                    fix: "Fix: CUDA resident dispatch launch resources were consumed before timing-event record.".to_string(),
                })?
                .timing_events()?
            {
                end_event.record(stream_raw)?;
            }
            // SAFETY: stream_raw is the live CUDA stream used for the launches
            // above. Native resident dispatch intentionally fences after the
            // kernel before host-visible output staging. The direct async DtoH/DtoD
            // path after a resident-staged launch can leave the completion event
            // unsignaled on current CUDA drivers, while an explicit post-kernel
            // fence followed by synchronous readback preserves correctness and
            // keeps the actual Program execution on CUDA instead of falling back
            // to host-buffer dispatch.
            crate::stream::synchronize_raw_stream(
                stream_raw,
                "cuStreamSynchronize (resident post-kernel)",
            )?;
            self.telemetry.record_sync_point();
            // Both inner events are complete: the fence above drained the
            // stream they were recorded on.
            if let Some((kernel_start, kernel_end)) = kernel_events.as_ref() {
                probe::record_kernel_ns(kernel_start.elapsed_time_ns(kernel_end)?);
            }
            if trace {
                tracing::debug!(
                    "[cuda-trace] +{}ms resident post-kernel sync complete",
                    start.elapsed().as_millis()
                );
            }
            Ok(())
        })();
        if let Err(error) = enqueue_result {
            let Some(launch_resources) = launch_resources.take() else {
                return Err(error);
            };
            match crate::stream::synchronize_raw_stream(
                stream_raw,
                "cuStreamSynchronize (resident async error cleanup)",
            ) {
                Ok(()) => {
                    self.telemetry.record_sync_point();
                    return Err(error);
                }
                Err(sync_error) => {
                    tracing::error!(
                        "Fix: failed to synchronize CUDA resident dispatch stream after enqueue error: {sync_error}. In-flight resident dispatch resources will not be recycled."
                    );
                    std::mem::forget(launch_resources);
                    if let Some(allocations) = allocations.take() {
                        std::mem::forget(allocations);
                    }
                    if let Some(resident_use) = resident_use.take() {
                        std::mem::forget(resident_use);
                    }
                    if let Some(host_transfers) = host_transfers.take() {
                        std::mem::forget(host_transfers);
                    }
                    return Err(error);
                }
            }
        }
        // Recycled only on the success path. An enqueue failure returns above
        // with the pair still owned here, and dropping it destroys the events
        // rather than returning them to a pool whose stream state is unproven.
        if let Some((kernel_start, kernel_end)) = kernel_events {
            self.launch_resources.release_timing_event(kernel_start);
            self.launch_resources.release_timing_event(kernel_end);
        }
        let launch_resources = launch_resources
            .take()
            .ok_or_else(|| BackendError::InvalidProgram {
                fix: "Fix: CUDA resident dispatch launch resources were consumed before synchronous output readback.".to_string(),
            })?;
        let allocations = allocations.take().ok_or_else(|| BackendError::InvalidProgram {
            fix: "Fix: CUDA resident dispatch allocations were consumed before synchronous output readback.".to_string(),
        })?;
        let resident_use = resident_use.take().ok_or_else(|| BackendError::InvalidProgram {
            fix: "Fix: CUDA resident dispatch use guard was consumed before synchronous output readback.".to_string(),
        })?;
        let mut host_transfers =
            host_transfers
                .take()
                .ok_or_else(|| BackendError::InvalidProgram {
                    fix: "Fix: CUDA resident dispatch host staging was consumed before synchronous output readback.".to_string(),
                })?;
        let mut staged_readback_bytes = 0_u64;
        let mut staged_readback_ops = 0_u64;
        // Stage every output with one stream-ordered async copy and a single
        // fence, rather than one blocking copy per output. A blocking copy
        // costs a full driver round trip per binding: measured on this box
        // (RTX 5090, driver 570.211.01), 16 blocking 16-byte readbacks take
        // 52.5 us against 17.8 us for 16 async copies plus one synchronize,
        // a 3.39x median over 8 ABBA rounds. The post-kernel fence above
        // already drained the stream, so these copies are enqueued onto an
        // idle stream and the fence below is what makes them host-visible.
        let staging_result = (|| -> Result<(), BackendError> {
            for &(src_base_ptr, readback) in &output_stage_readbacks {
                let dst = host_transfers.push_output(readback.byte_len)?;
                if readback.byte_len == 0 {
                    continue;
                }
                add_resident_dispatch_bytes(
                    &mut staged_readback_bytes,
                    readback.byte_len,
                    "resident staged output readback",
                )?;
                add_resident_dispatch_u64_count(
                    &mut staged_readback_ops,
                    "resident staged output readback operation",
                )?;
                let src_ptr = vyre_driver::accounting::checked_add_u64_usize_offset_lazy(
                    src_base_ptr,
                    readback.device_offset,
                    || {
                        BackendError::InvalidProgram {
                        fix: format!(
                            "Fix: CUDA resident staged output readback offset {} does not fit CUdeviceptr arithmetic.",
                            readback.device_offset
                        ),
                    }
                    },
                    || BackendError::InvalidProgram {
                        fix: format!(
                            "Fix: CUDA resident staged output pointer overflowed at offset {}.",
                            readback.device_offset
                        ),
                    },
                )?;
                // SAFETY: The source is the transient launch output buffer and
                // the destination is pinned host staging owned by
                // host_transfers, which stays alive (or is leaked below when
                // the fence fails) until every enqueued copy has retired.
                unsafe {
                    crate::backend::copy::d2h_async_checked_with_label(
                        dst,
                        src_ptr,
                        readback.byte_len,
                        stream_raw,
                        "cuMemcpyDtoHAsync_v2 (resident staged output)",
                    )?;
                }
            }
            Ok(())
        })();
        // The fence runs even when staging failed part way: copies already
        // enqueued are still in flight and their pinned destinations cannot be
        // recycled until the stream proves them retired.
        let fence_result = if staged_readback_ops == 0 {
            Ok(())
        } else {
            crate::stream::synchronize_raw_stream(
                stream_raw,
                "cuStreamSynchronize (resident staged output readback)",
            )
        };
        let staging_error = match (staging_result, fence_result) {
            (Ok(()), Ok(())) => None,
            (Err(error), Ok(())) => Some(error),
            (staging_outcome, Err(fence_error)) => {
                tracing::error!(
                    "Fix: failed to synchronize CUDA resident dispatch stream after staging output readbacks: {fence_error}. In-flight resident dispatch resources will not be recycled."
                );
                std::mem::forget(host_transfers);
                std::mem::forget(launch_resources);
                std::mem::forget(allocations);
                std::mem::forget(resident_use);
                return Err(staging_outcome.err().unwrap_or(fence_error));
            }
        };
        if let Some(error) = staging_error {
            return Err(error);
        }
        if staged_readback_ops != 0 {
            self.telemetry.record_sync_point();
        }
        self.telemetry
            .record_device_to_host_readback(staged_readback_bytes);
        self.telemetry
            .record_device_readback_operations(staged_readback_ops);
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident launch/output readbacks",
                start.elapsed().as_millis()
            );
        }
        let (stream, timing_events) = launch_resources.into_parts()?;
        let mut outputs = reserved_vec(output_stage_readbacks.len(), "resident staged output")?;
        host_transfers.collect_outputs_into(&mut outputs)?;
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident output collection complete",
                start.elapsed().as_millis()
            );
        }
        self.launch_resources.release_stream(stream);
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident stream released",
                start.elapsed().as_millis()
            );
        }
        let device_ns = match timing_events.as_ref() {
            Some((start_event, end_event)) => Some(start_event.elapsed_time_ns(end_event)?),
            None => None,
        };
        if let Some((start_event, end_event)) = timing_events {
            self.launch_resources.release_timing_event(start_event);
            self.launch_resources.release_timing_event(end_event);
        }
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident timing events released",
                start.elapsed().as_millis()
            );
        }
        drop(resident_use);
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident use released",
                start.elapsed().as_millis()
            );
        }
        drop(allocations);
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident allocations released",
                start.elapsed().as_millis()
            );
        }
        drop(host_transfers);
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident host transfers released",
                start.elapsed().as_millis()
            );
        }
        if trace {
            tracing::debug!(
                "[cuda-trace] +{}ms resident synchronous completion",
                start.elapsed().as_millis()
            );
        }
        Ok(CudaResidentDispatch {
            pending: crate::stream::CudaPendingDispatch::new_ready_timed(
                Arc::clone(&self.ctx),
                Arc::clone(&self.launch_resources),
                outputs,
                device_ns,
                Arc::clone(&self.telemetry),
            ),
            output_handles,
            output_readbacks,
        })
    }
}