gqa(adr): add supporting feature ADRs 0061-0063 for GQA fused attention

Proposed prerequisites surfaced while evaluating the GQA fused-attention ADR against the actual kernbench tl/sim_engine implementation: - ADR-0061 tl.broadcast: data-faithful GQA head reuse (fixes the MemoryStore nbytes check that forces h_q==h_kv==1 today). - ADR-0062 tl.load_async: non-blocking HBM tile load for KV prefetch (KV-load-bound decode/long-context overlap). - ADR-0063 tl.scratch_scope: per-tile scratch recycling (removes the 1 MiB bump-allocator ceiling that caps context at S=16). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-06-03 17:38:55 -07:00
parent 189794510a
commit f4f55b2c1c
3 changed files with 503 additions and 0 deletions
@@ -0,0 +1,193 @@
 # ADR-0061: `tl.broadcast` / `tl.repeat` — data-faithful GQA head reuse
 ## Status
 Proposed
 > Supporting ADR for **ADR-0060** (AHBM GQA Fused Attention). This ADR
 > removes the single hard blocker that keeps the existing attention
 > kernels at `h_q == h_kv == 1`: there is no way to expand a KV head to
 > the `G` query heads that share it without violating the MemoryStore
 > byte-conservation check.
 ## Context
 ### The GQA reuse lever
 Grouped-Query Attention amortises K/V loads: `G = H_q / H_kv` query heads
 share one KV head. For Llama3-70B `G = 8`. In a kernel this means a KV
 tile of shape `[S_kv, d]` (for one KV head) must be matmul'd against the
 `G` query rows of its group. Loading K/V **once** and reusing it across
 the group is the decode-time efficiency win (ADR-0060 §5.2).
 ### Why it does not work today
 The existing mesh kernels reshape tensors with a metadata-only helper
 `_view` (`src/kernbench/benches/_attention_mesh_kv.py:25-36`,
 `_attention_mesh_mlo.py:25-36`):
 ```python
 def _view(handle, new_shape):
    return TensorHandle(..., shape=new_shape, nbytes=handle.nbytes, ...)
    #                          ^^^ new shape          ^^^ OLD nbytes (unchanged)
 ```
 This rewrites `shape` but keeps the original `nbytes`. It is purely
 symbolic. It survives **Phase 1 (timing)** because Phase 1 never reads
 tensor bytes. It fails **Phase 2 (`enable_data=True`)** because the
 data executor materialises every op through `MemoryStore.read`, which
 enforces byte conservation
 (`src/kernbench/sim_engine/memory_store.py:67-73`):
 ```python
 if shape is not None and arr.shape != shape:
    if arr.nbytes != np.prod(shape) * arr.dtype.itemsize:
        raise ValueError(f"Shape mismatch: stored {arr.shape} ({arr.nbytes}B) "
                         f"vs requested {shape} ({...}B)")
    arr = arr.reshape(shape)
 ```
 A reshape conserves bytes; a **broadcast does not** (`[S_kv, 1, d]` →
 `[S_kv, G, d]` multiplies the element count by `G`). So the moment a
 kernel tries to view a `h_kv`-headed K as if it had `h_q` heads, the
 stored array has `1/G` of the requested bytes and `read` raises.
 The current test suite documents this as a deliberate limitation
 (`tests/attention/test_milestone_gqa_llama70b.py:137-142`):
 > "v1 uses `h_q == h_kv == 1` to avoid … GQA broadcast view (which is
 > symbolic and does not survive MemoryStore's nbytes check under
 > simulator data execution). Real GQA (`h_q > h_kv`) is deferred."
 There is **no** broadcast / repeat / expand primitive in the `tl` API
 today (`src/kernbench/triton_emu/tl_context.py` exposes `trans`,
 `zeros`, `full`, `arange` as metadata-only ops; none replicate data).
 ## Decision
 Add a **data-faithful broadcast op** to the `tl` API and the data
 executor. Unlike `_view`, it updates `nbytes` and, in Phase 2, actually
 replicates the underlying ndarray. The latency it charges models the
 on-chip register/SRAM fan-out, not an HBM re-load (the KV tile is read
 from HBM once; broadcast is an on-chip replication).
 ### D1. `tl` surface
 ```python
 def broadcast(self, x: TensorHandle, new_shape: tuple[int, ...]) -> TensorHandle:
    """Replicate x along size-1 axes to new_shape (numpy broadcast rules).
    new_shape must be broadcast-compatible with x.shape: every axis is
    either equal to x's, or x's is 1 (the axis being expanded). nbytes of
    the result reflect new_shape (unlike the metadata-only _view).
    """
 ```
 `tl.repeat(x, axis, n)` is offered as a thin convenience wrapper over
 `broadcast` for the common "insert a group axis and expand it to G"
 pattern, but `broadcast` is the primitive.
 ### D2. Command + op_log
 Emit a new `BroadcastCmd` (sibling of `MathCmd`):
 ```python
@dataclass(frozen=True)
 class BroadcastCmd:
    src: TensorHandle
    out: TensorHandle          # out.nbytes reflects new_shape
    new_shape: tuple[int, ...]
    data_op: bool = True
 ```
 op_log entry: `op_kind="math"`, `op_name="broadcast"` (it runs on the
 MATH/vector engine, like the other element-wise ops). The output handle
 is allocated from PE scratch via the existing `_make_compute_out`
 (`tl_context.py:148-156`), so `out.nbytes` is computed from `new_shape`
 and the byte-conservation check passes in Phase 2.
 ### D3. Phase 2 data execution
 Add `_execute_broadcast` to the data executor
 (`src/kernbench/sim_engine/data_executor.py`):
 ```python
 def _execute_broadcast(self, op):
    p = op.params
    src = self._resolve_read(p["src_space"], p["src_addr"],
                             p["src_shape"], p["dtype"], op.t_start)
    result = np.broadcast_to(src, p["new_shape"]).copy()   # .copy() → writable, owns bytes
    self.store.write(p["dst_space"], p["dst_addr"], result)
 ```
 `.copy()` is required: `np.broadcast_to` returns a zero-stride view, and
 MemoryStore stores references — a later consumer's nbytes check must see
 a real `prod(new_shape)`-element array.
 ### D4. Latency model
 Broadcast latency = vector-engine cost of writing `prod(new_shape)`
 elements (reuse `pe_math._compute_ns(num_elements)`,
 `src/kernbench/components/builtin/pe_math.py:46-51`). This models the
 on-chip replication into the register file / SRAM that real GQA hardware
 performs. It is **not** an HBM transaction — no `dma_read` is logged,
 consistent with "K/V tile loaded once" (ADR-0060 §5.2).
 ## Alternatives
 ### A1. Indexed per-head dot (no broadcast op)
 Loop the `G` query heads, each doing its own `tl.dot(Q_g, Kᵀ)` against
 the *same* (re-`ref`'d, not re-loaded) KV tile:
 ```python
 for g in range(G):
    s_g = tl.dot(Q[g], K_T)     # G separate GEMMs, KV reused via tl.ref
 ```
 - Pro: no new primitive; works in Phase 2 today.
 - Con: `G×` more GEMM commands (op_log inflation), no batched-GEMM
  modelling, and the per-call overhead is charged `G` times. It also
  obscures the "single batched matmul over the group" the hardware does.
  Rejected as the primary path; acceptable as a fallback if the batched
  2D dot's shape handling proves awkward.
 ### A2. Keep `_view`, relax the nbytes check
 Make `MemoryStore.read` tolerate broadcast (stride-0) shapes. Rejected:
 it silently breaks byte conservation for *all* readers and would mask
 genuine shape bugs elsewhere. The check is a correctness invariant
 (ADR-0052), not an obstacle to route around.
 ### A3. Materialise the broadcast at deploy time
 Store K/V already expanded to `h_q` heads in HBM. Rejected: defeats the
 entire point of GQA (it `G×`'s KV-cache HBM footprint and bandwidth).
 ## Consequences
 ### Positive
 - Unblocks real GQA (`h_q > h_kv`) in Phase 2 — the headline blocker for
  ADR-0060 and the deferred milestone work.
 - A general, reusable primitive (any kernel needing numpy-style
  broadcast benefits, not just attention).
 - Byte conservation stays intact; no weakening of MemoryStore.
 ### Negative
 - One new command type + one data-executor handler + one `tl` method.
 - Phase 2 `.copy()` allocates the expanded array in the store; for very
  large groups this is real memory. Bounded by tile size in the tiled
  design (ADR-0060), so acceptable.
 ## Test Requirements
 1. **Phase 1 timing**: `tl.broadcast([S,1,d] → [S,G,d])` logs one
   `math/broadcast` op with `num_elements == S*G*d`; no `dma_read`.
 2. **Phase 2 data**: broadcasting a known array yields the numpy
   `broadcast_to(...).copy()` result; a downstream `tl.dot` consuming it
   passes the MemoryStore nbytes check (the exact case that fails today).
 3. **GQA end-to-end**: re-run a reduced `milestone-gqa-llama70b` panel
   with `h_q=8, h_kv=1` under `enable_data=True`; it must complete (today
   it raises `ValueError: Shape mismatch`).
 4. **Broadcast-incompatible shape** (e.g. axis size 3 → 8) raises a
   clear error at emit time, not deep in Phase 2.
@@ -0,0 +1,148 @@
 # ADR-0062: `tl.load_async` — non-blocking HBM tile load for KV prefetch
 ## Status
 Proposed
 > Supporting ADR for **ADR-0060** (AHBM GQA Fused Attention). Decode and
 > long-context attention are **KV-load-bound**; the dominant lever is
 > overlapping the next K/V tile's DMA with the current tile's compute.
 > Today the only async primitive is for IPCQ comms, not for HBM loads.
 ## Context
 ### What overlap requires
 FlashAttention streams K/V tiles: while the MATH/GEMM engine works on
 tile `j` (Q·Kⱼᵀ → softmax → P·V), the DMA engine should already be
 pulling tile `j+1` (and `j+2` at prefetch depth 2). With that overlap a
 KV-load-bound kernel runs at roughly `max(compute, dma)` per tile instead
 of `compute + dma`.
 ### What exists
 - `tl.load(ptr, shape, dtype)` is **blocking**: it emits `DmaReadCmd`
  and the greenlet kernel suspends until PE_DMA signals completion
  (`src/kernbench/triton_emu/tl_context.py:177-203`; greenlet drive in
  `kernel_runner.py`). No two `tl.load`s can be in flight from one
  kernel.
 - An async pattern **does** exist, but only for IPCQ:
  `tl.recv_async(dir, ...) -> RecvFuture` + `tl.wait(future)`
  (`tl_context.py:543-560`, `tl_context.py:660-693`). It proves the
  machinery — a non-blocking command that returns a future, resolved
  later by `tl.wait` — works in the greenlet model.
 - The DMA engine already models a read channel as a SimPy resource
  separate from the write channel
  (`src/kernbench/components/builtin/pe_dma.py`), so concurrent in-flight
  reads are representable at the component layer; only the *kernel-facing
  API* serialises them.
 There is **no** `tl.load_async`. Prefetch/double-buffering cannot be
 expressed.
 ## Decision
 Add a non-blocking HBM load that mirrors the existing `recv_async`/`wait`
 contract.
 ### D1. `tl` surface
 ```python
 def load_async(self, ptr: int, shape: tuple[int, ...],
               dtype: str = "f16") -> LoadFuture:
    """Issue a DMA read and return immediately. Resolve with tl.wait(fut)
    -> TensorHandle. Multiple loads may be in flight; ordering of
    resolution is by tl.wait calls, not issue order."""
 ```
 `tl.wait` is extended to accept a `LoadFuture` (it already dispatches on
 `CompletionHandle` vs `RecvFuture` — add a third arm), returning the
 loaded `TensorHandle`.
 ### D2. Command
 Reuse the existing `DmaReadCmd` with a `blocking=False` flag (mirroring
 `IpcqRecvCmd.blocking`), or a thin `DmaReadAsyncCmd` sibling — whichever
 keeps PE_DMA's handler simplest. The op_log entry is unchanged
 (`memory/dma_read`); asynchrony is a scheduling property, not a new op
 kind, so existing op_log consumers and the `dma_read_count` metric
 (milestone bench) keep working.
 ### D3. Latency / overlap semantics
 - `load_async` charges the **issue** (descriptor push) only; the kernel
  proceeds.
 - The DMA transfer occupies the read channel for its modelled duration
  in parallel with whatever compute the kernel issues next.
 - `tl.wait(fut)` blocks only if the transfer has not finished; if it has,
  it returns immediately (same fast-path as `recv_async`/`wait`).
 - Determinism is preserved: completion is a scheduled event on the
  modelled DMA channel/link (SPEC §0.1, R8). No magic — the overlap is
  real modelled concurrency, not a hand-waved latency subtraction.
 ### D4. Double-buffer usage (the intended pattern)
 ```python
 # Prime: issue first two tile loads.
 f0 = tl.load_async(K_base + 0*tile_bytes, (TILE, d))
 f1 = tl.load_async(K_base + 1*tile_bytes, (TILE, d))
 for j in range(n_tiles):
    Kj = tl.wait(f_cur)                       # resolves the j-th load
    if j + 2 < n_tiles:                       # keep depth-2 pipeline full
        f_next2 = tl.load_async(K_base + (j+2)*tile_bytes, (TILE, d))
    s = tl.dot(q, tl.trans(Kj))               # compute overlaps f_{j+1}'s DMA
    ...
 ```
 (V tiles double-buffer the same way; ADR-0060 §3 schedules the V load so
 it overlaps the Q·Kᵀ + softmax of the same tile.)
 ## Alternatives
 ### A1. Use a tiled `tl.composite` instead
 `tl.composite` already pipelines tiles inside PE_SCHEDULER (DMA of tile
 `j+1` overlaps compute of tile `j`) and recycles per-tile scratch — so
 the composite path gets prefetch "for free." Rejected as the *sole*
 mechanism because today's composite is a single GEMM head + math
 epilogues (`pe_commands.py:144-162`); it cannot express the
 two-GEMM-with-running-`(m,l,O)` flash inner loop. Building a bespoke
 "flash composite" kind is a much larger change than `load_async`, which
 is a small, general primitive that also helps non-composite kernels.
 `load_async` and a future flash-composite are not mutually exclusive
 (ADR-0060 §8).
 ### A2. Rely on `recv_async` only (no HBM async)
 Only works for the Ring-Attention path where KV arrives over IPCQ
 (comms, already overlappable). The decode and no-SP paths load KV from
 **local HBM**, which has no async path today. Insufficient.
 ## Consequences
 ### Positive
 - Enables the headline KV-load-bound optimisation for decode and
  long-context, in the proven greenlet model.
 - Symmetric with `recv_async`/`wait` — low conceptual surface area.
 - General: any bandwidth-bound kernel can prefetch.
 ### Negative
 - Kernels must manage futures and buffer lifetimes explicitly (the
  double-buffer dance). Mitigated by keeping the pattern in a shared
  helper used by all four ADR-0060 cases.
 - Interacts with scratch lifetime (ADR-0063): in-flight buffers must not
  be recycled before their `tl.wait`. The recycling scope must exclude
  live prefetch buffers.
 ## Test Requirements
 1. **Overlap is real**: a kernel that issues `load_async` then a GEMM of
   comparable duration completes in ≈`max(load, gemm)`, not `load+gemm`
   (assert end-to-end latency strictly below the serial sum).
 2. **Correctness**: `tl.wait(load_async(...))` returns the same bytes as
   blocking `tl.load(...)` for the same address (Phase 2).
 3. **Two in flight**: two outstanding `load_async`s to distinct
   addresses both resolve to correct, independent tensors.
 4. **op_log compatibility**: each `load_async` still logs exactly one
   `memory/dma_read`, so `dma_read_count` metrics are unchanged vs the
   blocking version.
@@ -0,0 +1,162 @@
 # ADR-0063: `tl.scratch_scope` — per-tile scratch recycling for long context
 ## Status
 Proposed
 > Supporting ADR for **ADR-0060** (AHBM GQA Fused Attention). Long-context
 > attention sweeps many K/V tiles; each tile's intermediates
 > (`scores`, `P`, `exp`, partial `O`, …) currently allocate fresh
 > scratch that is never freed within a kernel invocation. The 1 MiB
 > per-PE scratch budget is exhausted long before a realistic context
 > length.
 ## Context
 ### The bump allocator
 `TLContext` allocates every math/compute output handle from a per-PE
 scratch pool with a **linear bump cursor**
 (`src/kernbench/triton_emu/tl_context.py`; `_scratch_alloc`):
 ```python
 def _scratch_alloc(self, nbytes):
    aligned = (nbytes + 15) & ~15
    addr = self._scratch_base + self._scratch_cursor
    self._scratch_cursor += aligned
    if self._scratch_cursor > self._scratch_size:      # default 1 << 20 = 1 MiB
        raise RuntimeError("TLContext scratch overflow: ...")
    return addr
 ```
 The docstring states the cursor **"resets on every kernel invocation"** —
 i.e. only at kernel entry, never *within* the kernel body. Every
 `tl.dot`, `tl.softmax`, `tl.exp`, `tl.sum`, `a - b`, `a * b`, … grabs a
 fresh slice and nothing is reclaimed until the kernel returns.
 ### Why this bites the GQA kernel
 The current milestone bench keeps `S_q = S_kv_per_rank = 16` **explicitly
 because** of this limit
 (`tests/attention/test_milestone_gqa_llama70b.py:123-148`):
 > "S_q_prefill and S_kv_per_rank are deliberately small (16 each) so the
 > simulator's 1 MB per-PE TCM kernel scratch is not exhausted by the
 > bump-allocated handle outputs of softmax/exp/dot/sum chains over
 > n_ranks ring steps."
 A FlashAttention sweep over `n_tiles` tiles allocates O(`n_tiles` ×
 per-tile-intermediates). For any realistic context this overflows. The
 *math* of flash attention needs only **O(1)** live scratch — the running
 `(m, l, O)` plus the current tile's working set — because each tile's
 temporaries are dead once that tile is folded into the running state. The
 allocator just doesn't know they're dead.
 ## Decision
 Add a **scratch scope** that lets a kernel mark a region of allocations
 as reclaimable, so per-tile temporaries are recycled while live running
 state (and in-flight prefetch buffers) are preserved.
 ### D1. `tl` surface — context manager
 ```python
 with tl.scratch_scope():
    s   = tl.dot(q, k_t)        # all handles allocated inside the `with`
    p   = tl.softmax(s)         # share a region that is rewound on exit
    o_j = tl.dot(p, v)
    # ... fold o_j into running (m,l,O) which live OUTSIDE the scope ...
 # on __exit__: cursor rewinds to its value at __enter__
 ```
 Semantics:
 - `__enter__` records the current `_scratch_cursor` as a save-point.
 - `__exit__` restores the cursor to the save-point, freeing everything
  allocated inside.
 - Handles allocated **outside** the scope (running `m,l,O`, prefetch
  buffers held by `LoadFuture` from ADR-0062) keep their addresses —
  they were allocated before the save-point or in an enclosing scope.
 ### D2. Safety contract
 A handle allocated inside a scope **must not** be read after the scope
 exits — its bytes may be overwritten by the next scope's allocations.
 The flash loop respects this naturally: the only values that survive a
 tile iteration are the running accumulators, which are allocated outside
 the per-tile scope and updated by ops *inside* it writing to outside
 addresses (the merge writes new running state — see D3).
 ### D3. Interaction with running accumulators
 The online-softmax merge reads the old running `(m, l, O)` and the
 current tile's `(m_j, l_j, O_j)`, producing new running values. To keep
 the new running values outside the recycled region, the merge writes them
 to **stable scratch** allocated once before the loop (a small fixed
 "running-state" arena, distinct from the per-tile scope). Concretely the
 kernel keeps two arenas:
 - **persistent arena** (allocated once): `m, l, O` (and their
  double-buffer if needed for the merge).
 - **scoped arena** (rewound each tile): `scores, P, exp, O_j, scale_*`.
 This mirrors how real flash-attention SRAM budgeting works: a small
 persistent accumulator region + a recycled tile working set.
 ### D4. Nesting
 Scopes nest (stack of save-points). Inner scope exit rewinds to the inner
 save-point; outer exit rewinds further. This supports an outer
 "per-query-block" scope around an inner "per-KV-tile" scope for prefill.
 ## Alternatives
 ### A1. Round-trip temporaries through HBM
 Store intermediates to HBM and reload to "free" TCM scratch. Rejected:
 turns a TCM-resident streaming kernel into an HBM-bandwidth-bound one —
 the opposite of the goal, and it pollutes op_log with spurious DMA.
 ### A2. Tiled `tl.composite` (scheduler-managed scratch)
 `tl.composite` recycles per-tile scratch inside PE_SCHEDULER
 automatically. As in ADR-0062 A1, this is attractive but blocked on a
 flash-capable composite kind (two GEMMs + carried `(m,l,O)`), a much
 larger change. `scratch_scope` gives the same memory behaviour for the
 greenlet kernel with a tiny, general primitive. The two can coexist.
 ### A3. Grow the scratch budget
 Bump `pe_tcm.kernel_scratch_mb`. Rejected as a fix: it only pushes the
 context-length ceiling out linearly while still leaking O(`n_tiles`)
 scratch, and it misrepresents the hardware (real TCM is small; the point
 of flash attention is O(1) working set). Useful only as a coarse knob,
 not a substitute for recycling.
 ## Consequences
 ### Positive
 - Removes the artificial `S = 16` validation-scale ceiling; enables
  realistic context lengths in both timing and data modes.
 - Faithfully models flash attention's O(1) working-set property.
 - Small, general primitive (any tiled kernel benefits).
 ### Negative
 - A use-after-scope bug silently reads stale bytes. Mitigated by the D2
  contract, by keeping the scope discipline inside a shared attention
  helper, and (optionally) by a debug build that poisons rewound regions.
 - Must coordinate with ADR-0062: prefetch buffers are live across tile
  iterations, so they belong to the persistent arena, not the scoped one.
 ## Test Requirements
 1. **Recycling**: a loop of `N` tiles inside `tl.scratch_scope()` keeps
   peak `_scratch_cursor` bounded by one tile's footprint, independent of
   `N` (today it grows linearly and overflows).
 2. **Correctness**: a flash sweep with scopes produces the same `O` as
   the same sweep without scopes at a small `N` that fits without
   recycling (Phase 2).
 3. **Long context**: a sweep at an `N` that would overflow 1 MiB without
   scopes completes (the exact failure the `S=16` cap avoids today).
 4. **Persistent-vs-scoped isolation**: running `(m,l,O)` allocated
   outside the scope retains correct values across `__exit__`.
 5. **Nesting**: nested scopes rewind to the correct save-points.