vllm.models.deepseek_v4.amd.mono.attention.back

K2 back: one attention layer from the rotated query to the rank's wo_b partial.

One launch a layer and rank, BLOCKS x THREADS, S rows:

split   (S x chunks of CK keys): flash-decoding over one chunk of a token's
        keys (K1's key table: top-k compressed rows, then window rows) for
        every head of the rank. The token's q in LDS, each head's RMSNorm
        factor (``q_head_rstd``) from it; each wave one
        (head tile, 16-key tile): its fp8_ds_mla rows loaded once,
        dequantized into its QK operands and kept raw in LDS; QK (bf16
        MFMA), the chunk's max / sum per head, bf16 p in LDS; PV over the
        chunk's raw rows (wave w: dims 64 w ..) -> the chunk's partial
        (bf16) and its (max, sum) per head, published
        A wide key table (``bounded``) skips the chunks past a token's
        keys, known from its position.
combine (S x H): a (token, head)'s chunks merged with the sink folded in
        (``bounded``: its live chunks only, ``merge_live``),
        bf16, inverse GPT-J RoPE of dims 448.., vLLM's MXFP8 of each 32
        -> XO (published)
wo_a    (G x 1024 / 32, two 16-row tiles): the group's XO rows times wo_a
        -> bf16 -> MXFP8 of the 32 rows -> X8B (published)
wo_b    (448, 16 rows): X8B times wo_b -> bf16 -> ``out``

Hand-offs are plain device-scope data behind tagged flags (half the bytes of a tagged pair per value). Adapted from ATOM's V4.1 attn_post (K2a), which reads a bf16 pool and ends in a peer all-reduce.

Functions:

• await_flags –

  Thread u < n polls flag index(u); then the CTA proceeds.

• back_context –

  The attention back's context: LDS views (lds: an allocated

• back_lds –

  K2's attention LDS: the union of back_lds_members.

• back_lds_members –

  K2's attention LDS structs: the split stage's and the GEMVs' (a union's

• back_plan –

  (nchunk, n_split, n_comb, COMB0, WOA0, WOB0, early_wob) of a build.

• chunk_keys –

  A split unit's keys: a 16-key tile a wave of a head group.

• epoch_begin –

  This CTA has read the launch pair's epoch: its own mark word (no

• epoch_end –

  CTA 0, once every CTA marked this epoch (thread u checks CTA u's mark):

• head_groups –

  Wave groups a split unit's head tiles are spread over: two (a tile each,

• key_rec –

  (data address, scale address, valid) of token t's key k from K1's key

• merge_all –

  The chunks' (sum, 8 dims) of (t, h) weighted by exp2(max - the max),

• merge_live –

  merge_all over the live chunks only, MERGE_BATCH at a time (a

• p_row –

  A P row (one head's chunk of bf16 p) in LDS, i32 words.

• publish –

  This CTA's stores visible device-wide, then the flag pair (region, i).

• q_head_rstd –

  Each head's weightless RMSNorm of the token's q in LDS:

• run_back –

  The attention back's stages on this CTA: split, combine, wo_a, wo_b.

• split_chunk –

  Flash-decoding over chunk ch of token t, every head of the rank

• stage_combine –

  (token, 8 heads) unit, a wave a head, 8 dims a lane: the chunks

• stage_split –

  Flash-decoding over chunk unit % nchunk of token unit // nchunk

• stage_woa –

  Rows 32 task .. of wo_a -> bf16 -> vLLM's MXFP8 of the 32 -> X8B.

• stage_wob –

  Rows 16 task .. (and 16 task2 .., has2) of wo_b -> bf16 -> out, or

_chunk_loads(c, t, h, d0, cc)

Chunk cc's (max, sum) of head h and its partial's 8 dims from d0.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def _chunk_loads(c, t, h, d0, cc):
    """Chunk cc's (max, sum) of head h and its partial's 8 dims from d0."""
    d, nchunk = c["d"], c["nchunk"]
    stv = fx.Vector(
        bo.buffer_load(
            rsrc(c["pstat"]),
            ((t * nchunk + cc) * d.heads + h) * 2,
            vec_width=2,
            dtype=T.f32,
        )
    )
    pv = fx.Vector(
        bo.buffer_load(
            rsrc(c["part"]),
            (((t * nchunk + cc) * d.heads + h) * HEAD_DIM + d0) // 2,
            vec_width=4,
            dtype=T.i32,
        )
    )
    vals = []
    for e in range_constexpr(4):
        vals += [bf_lo(pv[e]), bf_hi(pv[e])]
    return stv, vals

await_flags(c, region, n, index)

Thread u < n polls flag index(u); then the CTA proceeds.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def await_flags(c, region, n, index):
    """Thread u < n polls flag ``index(u)``; then the CTA proceeds."""
    tid = c["tid"]
    if tid < n:
        c["mb"].poll([(region, index(fx.min(tid, n - 1)), 1)])
    gpu.barrier()

back_context(key, lds, args, scratch, tag, layout=None)

The attention back's context: LDS views (lds: an allocated back_lds, or a dict of the members' views), mailbox (tag tag), scratch regions, arguments by name.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def back_context(key, lds, args: dict, scratch, tag, layout=None) -> dict:
    """The attention back's context: LDS views (``lds``: an allocated
    ``back_lds``, or a dict of the members' views), mailbox (tag ``tag``),
    scratch regions, arguments by name."""
    s, d = key.tokens, Dims(key.tp, key.topk)
    tid = fx.thread_idx.x
    if isinstance(lds, dict):
        spl, gml = lds["split"], lds["gemv"]
    else:
        spl, gml = lds.split.peek(), lds.gemv.peek()
    c = {
        "S": s,
        "tid": tid,
        "bid": fx.block_idx.x,
        "lane": tid % 64,
        "wave": tid // 64,
        "red": gml.red.ptr,
        "xl": gml.xl.ptr,
        "xsl": gml.xsl.ptr,
        "d": d,
        "ratio": key.ratio,
        "mb": Mailbox(tag),
        "nchunk": back_plan(key)[0],
        "sl": {
            "q": spl.q.ptr,
            "raw": spl.raw.ptr,
            "p": spl.p.ptr,
            "st": spl.st.ptr,
            "qn": spl.qn.ptr,
        },
        "args": args,
    }
    for region, (off, _) in (
        layout or back_scratch(s, d, start=front_end(s, d))
    ).items():
        c[region] = scratch + fx.Int64(off)
    return c

back_lds(s, d)

K2's attention LDS: the union of back_lds_members.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def back_lds(s: int, d: Dims):
    """K2's attention LDS: the union of ``back_lds_members``."""
    m = back_lds_members(s, d)
    SplitLds, GemvLds = m["split"], m["gemv"]

    @fx.union
    class BackLds:
        split: SplitLds  # type: ignore[valid-type]
        gemv: GemvLds  # type: ignore[valid-type]

    return BackLds

back_lds_members(s, d)

K2's attention LDS structs: the split stage's and the GEMVs' (a union's members: one at a time).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def back_lds_members(s: int, d: Dims) -> dict:
    """K2's attention LDS structs: the split stage's and the GEMVs' (a union's
    members: one at a time)."""
    rows = tile_rows(s)
    ck = chunk_keys(d, s)

    @fx.struct
    class SplitLds:
        q: fx.Array[fx.Int32, d.heads * Q_ROW, 16]
        raw: fx.Array[fx.Int32, ck * REC // 4, 16]
        p: fx.Array[fx.Int32, d.heads * p_row(d, s), 16]
        st: fx.Array[fx.Float32, 2 * d.heads * (WAVES // head_groups(d, s)), 16]
        # each head's q norm (``q_head_rstd``)
        qn: fx.Array[fx.Float32, d.heads, 16]

    @fx.struct
    class GemvLds:
        red: fx.Array[fx.Float32, 2 * RED, 16]
        xl: fx.Array[fx.Int32, rows * lds_row(d.group_k // 4), 16]
        xsl: fx.Array[fx.Int32, rows * lds_row(d.group_k // 32), 16]

    return {"split": SplitLds, "gemv": GemvLds}

back_plan(key)

(nchunk, n_split, n_comb, COMB0, WOA0, WOB0, early_wob) of a build.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def back_plan(key):
    """(nchunk, n_split, n_comb, COMB0, WOA0, WOB0, early_wob) of a build."""
    s, d = key.tokens, Dims(key.tp, key.topk)
    ck = chunk_keys(d, s)
    nchunk = cdiv(d.keys, ck)
    n_split = s * nchunk
    n_comb = s * (d.heads // COMB_HEADS)
    COMB0 = n_split % BLOCKS
    assert d.woa_tasks <= BLOCKS
    assert BLOCKS < WOB_TASKS <= 2 * BLOCKS
    # the GEMVs past the attention units, so CTAs without attention work have
    # their weights in flight from the start
    WOA0 = (n_split + n_comb) % BLOCKS
    WOB0 = (WOA0 + d.woa_tasks) % BLOCKS
    # wo_b's weights before the wo_a stage only while the wo_a CTAs have no
    # attention work (loads complete in issue order: else they would hold up
    # wo_a's activation rows)
    early_wob = n_split + n_comb + d.woa_tasks <= BLOCKS
    return nchunk, n_split, n_comb, COMB0, WOA0, WOB0, early_wob

chunk_keys(d, s)

A split unit's keys: a 16-key tile a wave of a head group.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def chunk_keys(d: Dims, s: int) -> int:
    """A split unit's keys: a 16-key tile a wave of a head group."""
    return 16 * (WAVES // head_groups(d, s))

epoch_begin(c, epoch, ep)

This CTA has read the launch pair's epoch: its own mark word (no contended atomic) := ep + 1, which epoch_end checks.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def epoch_begin(c, epoch, ep):
    """This CTA has read the launch pair's epoch: its own mark word (no
    contended atomic) := ep + 1, which ``epoch_end`` checks."""
    if c["tid"] == 0:
        bo.buffer_store(
            ep + 1, rsrc(epoch), EPOCH_MARKS + c["bid"], cache_modifier=CM_DEV
        )

epoch_end(c, epoch, ep, reset=None)

CTA 0, once every CTA marked this epoch (thread u checks CTA u's mark): reset() (state later pairs must find cleared), then the epoch moved on. The next pair's hand-offs carry a tag no earlier pair wrote; later launches' kernels start after this one ends.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def epoch_end(c, epoch, ep, reset=None):
    """CTA 0, once every CTA marked this epoch (thread u checks CTA u's mark):
    ``reset()`` (state later pairs must find cleared), then the epoch moved on.
    The next pair's hand-offs carry a tag no earlier pair wrote; later launches'
    kernels start after this one ends."""
    if c["bid"] == 0:
        tid = c["tid"]
        if tid < BLOCKS:
            r = rsrc(epoch)

            def load_mark():
                return fx.Int32(
                    bo.buffer_load(
                        r,
                        EPOCH_MARKS + tid,
                        vec_width=1,
                        dtype=T.i32,
                        cache_modifier=CM_DEV,
                    )
                )

            _ = spin_until(load_mark, lambda v: v != ep + 1)
        gpu.barrier()
        if tid == 0:
            if const_expr(reset is not None):
                reset()
            gstore(epoch, ep + 1, words=1)

head_groups(d, s)

Wave groups a split unit's head tiles are spread over: two (a tile each, 64-key chunks, twice the units) while the units fit the grid, else one (a wave's keys loaded and dequantized once for every head tile).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def head_groups(d: Dims, s: int) -> int:
    """Wave groups a split unit's head tiles are spread over: two (a tile each,
    64-key chunks, twice the units) while the units fit the grid, else one (a
    wave's keys loaded and dequantized once for every head tile)."""
    return 2 if d.head_tiles == 2 and s * cdiv(d.keys, 64) <= BLOCKS else 1

key_rec(c, t, k)

(data address, scale address, valid) of token t's key k from K1's key table: slot | 1 << 31 for a compressed row, a window slot otherwise, -1 for none (it reads a zero record: its scores are masked, its values zeros).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def key_rec(c, t, k):
    """(data address, scale address, valid) of token t's key k from K1's key
    table: slot | 1 << 31 for a compressed row, a window slot otherwise, -1 for
    none (it reads a zero record: its scores are masked, its values zeros)."""
    a, ratio, keys = c["args"], c["ratio"], c["d"].keys
    # a table this launch wrote (``krdy``) is read past the CU's cache
    cm = CM_DEV if const_expr("krdy" in c) else 0
    val = ld_i32(a["kt"], t * keys + fx.min(fx.max(k, 0), keys - 1), cm)
    ok = val != fx.Int32(-1)
    slot = ok.select(val & fx.Int32(0x7FFFFFFF), fx.Int32(0))
    if const_expr(ratio > 0):
        is_c = ok & (val < 0)
        blk = is_c.select(a["comp_block"], a["swa_block"])
        stride = is_c.select(a["comp_stride"], a["swa_stride"])
        base = is_c.select(a["comp"], a["swa"])
    else:
        blk, stride, base = a["swa_block"], a["swa_stride"], a["swa"]
    b = slot // blk
    off = slot % blk
    page = base + fx.Int64(b) * stride
    z = a["zrec"]
    data = ok.select(page + fx.Int64(off * DATA), z)
    sc = ok.select(page + fx.Int64(blk * DATA + off * 8), z + fx.Int64(DATA))
    return data, sc, ok

merge_all(c, t, h, d0, nlive, sink)

The chunks' (sum, 8 dims) of (t, h) weighted by exp2(max - the max), the sink's term first, in chunk order: every chunk loaded at once (a past-nlive chunk weighs nothing).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def merge_all(c, t, h, d0, nlive, sink):
    """The chunks' (sum, 8 dims) of (t, h) weighted by exp2(max - the max),
    the sink's term first, in chunk order: every chunk loaded at once (a
    past-``nlive`` chunk weighs nothing)."""
    m = fx.Float32(NEG)
    ms, ls, vs = [], [], []
    for cc in range_constexpr(c["nchunk"]):
        live = cc < nlive
        stv, vals = _chunk_loads(c, t, h, d0, cc)
        mc = fx.Float32(live.select(stv[0], fx.Float32(NEG)))
        ms.append(mc)
        ls.append(fx.Float32(live.select(stv[1], fx.Float32(0.0))))
        vs.append([fx.Float32(live.select(v, fx.Float32(0.0))) for v in vals])
        m = fx.max(m, mc)
    m_final = fx.max(m, sink)
    lsum = hw_exp2(sink - m_final)
    tot = [fx.Float32(0.0)] * 8
    for cc in range_constexpr(c["nchunk"]):
        wgt = hw_exp2(ms[cc] - m_final)
        lsum = lsum + ls[cc] * wgt
        tot = [tot[e] + vs[cc][e] * wgt for e in range(8)]
    return lsum, tot

merge_live(c, t, h, d0, nlive, sink)

merge_all over the live chunks only, MERGE_BATCH at a time (a wide key table: most of its chunks past a token's keys).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def merge_live(c, t, h, d0, nlive, sink):
    """``merge_all`` over the live chunks only, ``MERGE_BATCH`` at a time (a
    wide key table: most of its chunks past a token's keys)."""
    m = fx.Float32(NEG)
    for c0 in range(0, nlive, MERGE_BATCH):
        for u in range_constexpr(MERGE_BATCH):
            live = c0 + u < nlive
            stv, _ = _chunk_loads(c, t, h, d0, fx.min(c0 + u, nlive - 1))
            m = fx.max(m, fx.Float32(live.select(stv[0], fx.Float32(NEG))))
    m_final = fx.max(m, sink)
    lsum = hw_exp2(sink - m_final)
    t0, t1, t2, t3, t4, t5, t6, t7 = [fx.Float32(0.0)] * 8
    for c0 in range(0, nlive, MERGE_BATCH):
        got = [
            _chunk_loads(c, t, h, d0, fx.min(c0 + u, nlive - 1))
            for u in range(MERGE_BATCH)
        ]
        for u in range_constexpr(MERGE_BATCH):
            live = c0 + u < nlive
            stv, vals = got[u]
            wgt = hw_exp2(fx.Float32(live.select(stv[0], fx.Float32(NEG))) - m_final)
            lsum = lsum + fx.Float32(live.select(stv[1], fx.Float32(0.0))) * wgt
            v = [fx.Float32(live.select(x, fx.Float32(0.0))) for x in vals]
            t0, t1, t2, t3 = (
                t0 + v[0] * wgt,
                t1 + v[1] * wgt,
                t2 + v[2] * wgt,
                t3 + v[3] * wgt,
            )
            t4, t5, t6, t7 = (
                t4 + v[4] * wgt,
                t5 + v[5] * wgt,
                t6 + v[6] * wgt,
                t7 + v[7] * wgt,
            )
    return lsum, [t0, t1, t2, t3, t4, t5, t6, t7]

p_row(d, s)

A P row (one head's chunk of bf16 p) in LDS, i32 words.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

def p_row(d: Dims, s: int) -> int:
    """A P row (one head's chunk of bf16 p) in LDS, i32 words."""
    return chunk_keys(d, s) // 2 + P_PAD

publish(c, region, i)

This CTA's stores visible device-wide, then the flag pair (region, i).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def publish(c, region, i):
    """This CTA's stores visible device-wide, then the flag pair (region, i)."""
    rocdl.s_waitcnt(vmcnt=0)
    gpu.barrier()
    if c["tid"] == 0:
        c["mb"].put(region, i, fx.Int32(1))

q_head_rstd(c)

Each head's weightless RMSNorm of the token's q in LDS: rsq(mean(q^2) + eps) -> qn [head]. vLLM norms q before its RoPE, which keeps every pair's norm: the scores scale by it after the dot product.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def q_head_rstd(c):
    """Each head's weightless RMSNorm of the token's q in LDS:
    rsq(mean(q^2) + eps) -> ``qn`` [head]. vLLM norms q before its RoPE, which
    keeps every pair's norm: the scores scale by it after the dot product."""
    tid, d, sl = c["tid"], c["d"], c["sl"]
    tph = THREADS // d.heads  # threads a head
    words = HEAD_DIM // 2 // tph
    h = tid // tph
    i = tid % tph
    acc = fx.Float32(0.0)
    for w in range_constexpr(words):
        v = fx.Int32(fx.ptr_load(sl["q"] + (h * Q_ROW + i * words + w)))
        lo, hi = bf_lo(v), bf_hi(v)
        acc = acc + lo * lo + hi * hi
    offs = tuple(o for o in (32, 16, 8, 4, 2, 1) if o < tph)
    acc = butterfly(acc, offs)
    if i == 0:
        mean = acc * fx.Float32(1.0 / HEAD_DIM)
        fx.ptr_store(hw_rsq(mean + fx.Float32(NORM_EPS)), sl["qn"] + h)
    gpu.barrier()

run_back(c, key, bid)

The attention back's stages on this CTA: split, combine, wo_a, wo_b.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def run_back(c, key, bid):
    """The attention back's stages on this CTA: split, combine, wo_a, wo_b."""
    s = key.tokens
    dd = Dims(key.tp, key.topk)
    # plain ints: a Python object read inside traced control flow is carried
    woa_tasks, groups, heads, group_k = dd.woa_tasks, dd.groups, dd.heads, dd.group_k
    q_rows, o_rows, wob_tasks = dd.q_rows, dd.o_rows, WOB_TASKS
    a = c["args"]
    tiles = token_tiles(s)
    _, n_split, n_comb, COMB0, WOA0, WOB0, early_wob = back_plan(key)
    stamp(c, 0)
    for unit in range(first_task(bid, 0), n_split, BLOCKS):
        stage_split(c, unit)
    stamp(c, 1)
    for unit in range(first_task(bid, COMB0), n_comb, BLOCKS):
        stage_combine(c, unit)
    gpu.barrier()
    stamp(c, 2)
    # the GEMV weights, every CTA right after its attention work: its wo_a
    # task (if any), then its one or two wo_b tasks
    woa_task = first_task(bid, WOA0)
    has_woa = woa_task < woa_tasks
    woa_t = fx.min(woa_task, woa_tasks - 1)
    ops = gemv_loads(
        c,
        a["woa"],
        a["woa_s"],
        group_k,
        woa_t * WOA_ROWS,
        WOA_SPLIT_K,
        tiled=True,
        pred=has_woa,
        ld=ld(group_k),
    )
    wob_task = first_task(bid, WOB0)
    wob_second = wob_task + BLOCKS
    has_second = wob_second < wob_tasks

    def wob_weights():
        return (
            wob_loads(c, wob_task),
            wob_loads(c, fx.min(wob_second, wob_tasks - 1), pred=has_second),
        )

    if const_expr(early_wob):
        ops1, ops2 = wob_weights()
    if const_expr(c.get("tl") is not None):
        rocdl.s_waitcnt(vmcnt=0)
        gpu.barrier()
    stamp(c, 8)
    # wo_a: every token tile of its group's XO rows, the next tile's in flight
    if has_woa:
        grp = woa_t // (woa_tasks // groups)
        await_flags(c, c["crdy"], s, lambda u: u * (heads // COMB_HEADS) + grp)
        stamp(c, 9)
        xo = (
            c["xo"],
            c["xos"],
            q_rows // 4,
            grp * group_k // 4,
            group_k // 4,
            q_rows // 32,
            grp * group_k // 32,
            group_k // 32,
        )
        regs = rows_loads(c, *xo, *tiles[0])
        for ti in range_constexpr(len(tiles)):
            t0, n = tiles[ti]
            rows_store(c, regs, group_k // 4, group_k // 32, c["xl"], c["xsl"], n)
            if const_expr(ti + 1 < len(tiles)):
                regs = rows_loads(c, *xo, *tiles[ti + 1])
            stage_woa(c, woa_t, ops, t0, n)
        stamp(c, 10)
        publish(c, c["ardy"], woa_t)
    stamp(c, 3)
    # wo_b
    if const_expr(not early_wob):
        ops1, ops2 = wob_weights()
    stamp(c, 11)
    await_flags(c, c["ardy"], woa_tasks, lambda u: u)
    stamp(c, 12)
    x8b = (
        c["x8b"],
        c["x8bs"],
        o_rows // 4,
        0,
        o_rows // 4,
        o_rows // 32,
        0,
        o_rows // 32,
    )
    regs = rows_loads(c, *x8b, *tiles[0])
    for ti in range_constexpr(len(tiles)):
        t0, n = tiles[ti]
        rows_store(c, regs, o_rows // 4, o_rows // 32, c["xl"], c["xsl"], n)
        if const_expr(ti == 0):
            stamp(c, 14)
        if const_expr(ti + 1 < len(tiles)):
            regs = rows_loads(c, *x8b, *tiles[ti + 1])
        stage_wob(
            c,
            wob_task,
            ops1,
            fx.min(wob_second, wob_tasks - 1),
            ops2,
            has_second,
            t0,
            n,
            mark=ti == 0,
        )
        if const_expr(ti == 0):
            stamp(c, 13)
    stamp(c, 4)

split_chunk(c, t, ch)

Flash-decoding over chunk ch of token t, every head of the rank -> the chunk's partial and (max, sum) per head. Wave (head group hg, key tile kt): keys 16 kt .. of the chunk (loaded and dequantized once) against the group's head tiles; then dims 64 w .. of the PV over every key and head.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def split_chunk(c, t, ch):
    """Flash-decoding over chunk ``ch`` of token ``t``, every head of the rank
    -> the chunk's partial and (max, sum) per head. Wave (head group hg, key
    tile kt): keys 16 kt .. of the chunk (loaded and dequantized once) against
    the group's head tiles; then dims 64 w .. of the PV over every key and
    head."""
    tid, lane, wave, d, s = c["tid"], c["lane"], c["wave"], c["d"], c["S"]
    a, sl = c["args"], c["sl"]
    nchunk = c["nchunk"]
    ck = chunk_keys(d, s)
    ngrp = head_groups(d, s)
    nkt = WAVES // ngrp  # 16-key tiles a chunk
    nht = d.head_tiles // ngrp  # head tiles a wave
    hgi = wave // nkt
    kt = wave % nkt
    g = lane // 16
    r16 = lane % 16
    k0 = ch * ck
    if const_expr("krdy" in c):
        await_flags(c, c["krdy"], 1, lambda u: t)
    # the key count, this lane's key and q: independent loads, out together
    # (the key table holds -1 past a token's keys)
    kv_len = ld_i32(a["klen"], t, CM_DEV if const_expr("krdy" in c) else 0)
    key = k0 + kt * 16 + r16
    data, sc, okv = key_rec(c, t, key)
    qrow = d.heads * HEAD_DIM // 8
    qv = []
    for i in range_constexpr(cdiv(qrow, THREADS)):
        u = fresh(tid) + THREADS * i
        qv.append(
            fx.Vector(
                bo.buffer_load(
                    rsrc(a["q"]),
                    (t * d.q_rows) // 2 + fx.min(u, qrow - 1) * 4,
                    vec_width=4,
                    dtype=T.i32,
                )
            )
        )
    if k0 < kv_len:
        # a key's 448 e4m3 as 7 x 16 B (lane g: dims 64 c + 16 g ..), its 64
        # bf16 rope dims as 2 x 16 B (32 c + 8 g ..), its 7 scale bytes
        scw = gload(sc, words=2)
        nope = [
            gload(data + fx.Int64(64 * c4 + 16 * g), words=4)
            for c4 in range(NOPE // 64)
        ]
        rope = [
            gload(data + fx.Int64(NOPE + (cc * 32 + 8 * g) * 2), words=4)
            for cc in range(ROPE // 32)
        ]
        # q of every head into LDS (bf16 rows of 512 + pad), 16 B a thread
        for i in range_constexpr(cdiv(qrow, THREADS)):
            u = fresh(tid) + THREADS * i
            if u < qrow:
                fx.ptr_store(
                    qv[i],
                    sl["q"]
                    + (u // (HEAD_DIM // 8)) * Q_ROW
                    + (u % (HEAD_DIM // 8)) * 4,
                )
        # the raw rows to LDS for the PV (head group 0's waves)
        if hgi == 0:
            rawb = sl["raw"] + (kt * 16 + r16) * (REC // 4)
            for c4 in range_constexpr(NOPE // 64):
                fx.ptr_store(nope[c4], rawb + (64 * c4 + 16 * g) // 4)
            for cc in range_constexpr(ROPE // 32):
                fx.ptr_store(rope[cc], rawb + (NOPE + (cc * 32 + 8 * g) * 2) // 4)
            if g == 0:
                fx.ptr_store(scw, rawb + DATA // 4)
        gpu.barrier()
        q_head_rstd(c)
        stamp(c, 5)
        # QK: A = the keys (dequantized once), B = q of a head tile (LDS); the
        # contraction's 32-dim chunk 2 c + h of the nope dims is {64 c + 16 g +
        # 8 h ..} (lane g's 16 B load), q read in the same order
        scale = fx.Int32(a["qk_scale"]).bitcast(fx.Float32)
        st = sl["st"]
        accs = [fx.Vector.filled(4, 0.0, fx.Float32) for _ in range(nht)]
        for c4 in range_constexpr(NOPE // 64):
            ksc = pow2(_scale_byte(scw, c4))
            for h2 in range_constexpr(2):
                kw = fp8x8_bf16(nope[c4][2 * h2], nope[c4][2 * h2 + 1], ksc).bitcast(
                    fx.BFloat16
                )
                for ht in range_constexpr(nht):
                    qb = sl["q"] + ((hgi * nht + ht) * HEAD_TILE + r16) * Q_ROW
                    qw = fx.Vector(
                        fx.ptr_load(
                            qb + 32 * c4 + 8 * g + 4 * h2,
                            result_type=fx.Vector.make_type(4, fx.Int32),
                        )
                    )
                    accs[ht] = mfma_bf16(kw, qw.bitcast(fx.BFloat16), accs[ht])
        for cc in range_constexpr(ROPE // 32):
            kw = rope[cc].bitcast(fx.BFloat16)
            for ht in range_constexpr(nht):
                qb = sl["q"] + ((hgi * nht + ht) * HEAD_TILE + r16) * Q_ROW
                qw = fx.Vector(
                    fx.ptr_load(
                        qb + (NOPE + cc * 32 + 8 * g) // 2,
                        result_type=fx.Vector.make_type(4, fx.Int32),
                    )
                )
                accs[ht] = mfma_bf16(kw, qw.bitcast(fx.BFloat16), accs[ht])
        # scores of keys 4 g + i (C rows), head r16 of a tile (C column)
        ok = (key < kv_len) & okv
        okw = ok.select(fx.Int32(1), fx.Int32(0))
        valid = [lane_gather(okw, 4 * g + i) != 0 for i in range(4)]
        svs = []
        for ht in range_constexpr(nht):
            head = (hgi * nht + ht) * HEAD_TILE + r16
            hscale = scale * fx.ptr_load(sl["qn"] + head)
            sv = [
                fx.Float32(valid[i].select(accs[ht][i] * hscale, fx.Float32(NEG)))
                for i in range(4)
            ]
            mw = butterfly(
                fx.max(fx.max(sv[0], sv[1]), fx.max(sv[2], sv[3])), (32, 16), fx.max
            )
            if g == 0:
                fx.ptr_store(mw, st + ((hgi * nht + ht) * nkt + kt) * HEAD_TILE + r16)
            svs.append(sv)
        gpu.barrier()
        lst = st + d.heads * nkt
        for ht in range_constexpr(nht):
            hti = hgi * nht + ht
            m = fx.ptr_load(st + (hti * nkt) * HEAD_TILE + r16)
            for j in range_constexpr(1, nkt):
                m = fx.max(m, fx.ptr_load(st + (hti * nkt + j) * HEAD_TILE + r16))
            sv = svs[ht]
            p = [
                fx.Float32(valid[i].select(hw_exp2(sv[i] - m), fx.Float32(0.0)))
                for i in range(4)
            ]
            lw = (p[0] + p[2]) + (p[1] + p[3])
            lw = lw + xshfl(lw, 32)
            lw = lw + xshfl(lw, 16)
            pw = (
                fx.Vector.from_elements(p, fx.Float32).to(fx.BFloat16).bitcast(fx.Int32)
            )
            pbase = (
                sl["p"] + (hti * HEAD_TILE + r16) * p_row(d, s) + (kt * 16 + 4 * g) // 2
            )
            fx.ptr_store(fx.Vector.from_elements([pw[0], pw[1]], fx.Int32), pbase)
            if g == 0:
                fx.ptr_store(lw, lst + (hti * nkt + kt) * HEAD_TILE + r16)
        gpu.barrier()
        # chunk stats: (max, sum) per head
        if tid < d.heads:
            hts = tid // HEAD_TILE
            hr = tid % HEAD_TILE
            mm = fx.ptr_load(st + (hts * nkt) * HEAD_TILE + hr)
            ll = fx.ptr_load(lst + (hts * nkt) * HEAD_TILE + hr)
            for j in range_constexpr(1, nkt):
                mm = fx.max(mm, fx.ptr_load(st + (hts * nkt + j) * HEAD_TILE + hr))
                ll = ll + fx.ptr_load(lst + (hts * nkt + j) * HEAD_TILE + hr)
            bo.buffer_store(
                fx.Vector.from_elements([mm, ll], fx.Float32),
                rsrc(c["pstat"]),
                ((t * nchunk + ch) * d.heads + tid) * 2,
                cache_modifier=CM_DEV,
            )
        # PV: wave w dims 64 w + 4 r16 + dt (dt = 0..3: one LDS word holds the
        # four of a key); the rope wave (dims 448..) reads bf16 pairs. Lane
        # (r16, g) takes keys 4 g + j of a 16-key tile: A[dim][key]
        is_rope = wave == WAVES - 1
        dbase = 64 * wave
        sh = (wave % 4) * 8
        accs = [
            [fx.Vector.filled(4, 0.0, fx.Float32) for _ in range(4)]
            for _ in range(d.head_tiles)
        ]
        for kt2 in range_constexpr(nkt):
            rb0 = sl["raw"] + (kt2 * 16 + 4 * g) * (REC // 4)
            vv = fx.Vector.filled(16, 0.0, fx.Float32)  # [dt][j]
            if is_rope:
                rv = []
                for j in range_constexpr(4):
                    rw = fx.Vector(
                        fx.ptr_load(
                            rb0 + j * (REC // 4) + (NOPE + 8 * r16) // 4,
                            result_type=fx.Vector.make_type(2, fx.Int32),
                        )
                    )
                    rv.append([bf_lo(rw[0]), bf_hi(rw[0]), bf_lo(rw[1]), bf_hi(rw[1])])
                vv = fx.Vector.from_elements(
                    [rv[j][dt] for dt in range(4) for j in range(4)], fx.Float32
                )
            else:
                nv = []
                for j in range_constexpr(4):
                    rb = rb0 + j * (REC // 4)
                    w = fx.Int32(fx.ptr_load(rb + (dbase + 4 * r16) // 4))
                    scb = fx.Int32(fx.ptr_load(rb + DATA // 4 + wave // 4))
                    scl = pow2((scb >> sh) & 0xFF)
                    nv.append(
                        _fp8_pair_f32(w, scl, False) + _fp8_pair_f32(w, scl, True)
                    )
                vv = fx.Vector.from_elements(
                    [nv[j][dt] for dt in range(4) for j in range(4)], fx.Float32
                )
            for ht2 in range_constexpr(d.head_tiles):
                pb = fx.Vector(
                    fx.ptr_load(
                        sl["p"]
                        + (ht2 * HEAD_TILE + r16) * p_row(d, s)
                        + (kt2 * 16 + 4 * g) // 2,
                        result_type=fx.Vector.make_type(2, fx.Int32),
                    )
                ).bitcast(fx.Int16)
                for dt in range_constexpr(4):
                    av = (
                        fx.Vector.from_elements(
                            [vv[dt * 4 + j] for j in range(4)], fx.Float32
                        )
                        .to(fx.BFloat16)
                        .bitcast(fx.Int16)
                    )
                    accs[ht2][dt] = fx.Vector(
                        rocdl.mfma_f32_16x16x16bf16_1k(
                            T.vec(4, T.f32), [av, pb, accs[ht2][dt]]
                        )
                    )
        stamp(c, 6)
        # partial[t][ch][head][dim] bf16: lane (r16, g) holds C rows 4 g + i ->
        # dims 64 w + 16 g + 4 i + dt of head ht2 * 16 + r16: 32 contiguous
        # bytes, two 16 B stores (4 lanes fill a 128 B line)
        for ht2 in range_constexpr(d.head_tiles):
            hh = ht2 * HEAD_TILE + r16
            words = []
            for i in range_constexpr(4):
                v4 = fx.Vector.from_elements(
                    [accs[ht2][dt][i] for dt in range(4)], fx.Float32
                )
                w2 = v4.to(fx.BFloat16).bitcast(fx.Int32)
                words += [w2[0], w2[1]]
            off = (((t * nchunk + ch) * d.heads + hh) * HEAD_DIM + dbase + 16 * g) // 2
            for hlf in range_constexpr(2):
                bo.buffer_store(
                    fx.Vector.from_elements(words[4 * hlf : 4 * hlf + 4], fx.Int32),
                    rsrc(c["part"]),
                    off + 4 * hlf,
                    cache_modifier=CM_DEV,
                )
        publish(c, c["srdy"], t * nchunk + ch)
        stamp(c, 7)

stage_combine(c, unit)

(token, 8 heads) unit, a wave a head, 8 dims a lane: the chunks merged with the sink folded in, bf16, inverse RoPE, MXFP8 of each 32 -> XO, published (a flag a (token, head group)).

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def stage_combine(c, unit):
    """(token, 8 heads) ``unit``, a wave a head, 8 dims a lane: the chunks
    merged with the sink folded in, bf16, inverse RoPE, MXFP8 of each 32 -> XO,
    published (a flag a (token, head group))."""
    lane, wave, d, s = c["lane"], c["wave"], c["d"], c["S"]
    a = c["args"]
    hg_n = d.heads // COMB_HEADS
    t = unit // hg_n
    hg = unit % hg_n
    h = hg * COMB_HEADS + wave
    nchunk = c["nchunk"]
    ck = chunk_keys(d, s)
    if const_expr("krdy" in c):
        # a key table this launch builds: its count once published, past the
        # CU's cache (a CTA without split work gets here first)
        await_flags(c, c["krdy"], 1, lambda u: t)
    kv_len = ld_i32(a["klen"], t, CM_DEV if const_expr("krdy" in c) else 0)
    nlive = (kv_len + ck - 1) // ck
    pos = ld_i32(a["pos"], 2 * t)
    sink = ld_f32(a["sink"], h) * fx.Float32(LOG2E)
    d0 = lane * 8  # this lane's 8 dims
    rope_lane = d0 >= NOPE
    jj = fx.max((d0 - NOPE) // 2, 0)  # 4 pairs from here
    cs = [
        rope_lane.select(ld_f32(a["cos_sin"], pos * ROPE + jj + i), fx.Float32(1.0))
        for i in range(4)
    ]
    sn = [
        rope_lane.select(
            ld_f32(a["cos_sin"], pos * ROPE + HALF + jj + i), fx.Float32(0.0)
        )
        for i in range(4)
    ]
    await_flags(c, c["srdy"], nlive, lambda u: t * nchunk + u)
    if const_expr(c.get("bounded", False)):
        lsum, tot = merge_live(c, t, h, d0, nlive, sink)
    else:
        lsum, tot = merge_all(c, t, h, d0, nlive, sink)
    inv = fx.Float32(1.0) / lsum
    o = [bf16_round(tot[e] * inv) for e in range(8)]
    # inverse GPT-J RoPE of each pair: e' = e cos + o sin, o' = o cos - e sin
    y = []
    for i in range_constexpr(4):
        e, od = o[2 * i], o[2 * i + 1]
        y += [e * cs[i] + od * sn[i], od * cs[i] - e * sn[i]]
    amax = abs(y[0])
    for v in y[1:]:
        amax = fx.max(amax, abs(v))
    amax = butterfly(amax, (1, 2), fx.max)  # 32 dims = 4 lanes
    code = mx_code(amax)
    mul = mx_mul(code)
    w0 = fp8_pack4(*[clamp_fp8(y[i] * mul) for i in range(4)])
    w1 = fp8_pack4(*[clamp_fp8(y[4 + i] * mul) for i in range(4)])
    bo.buffer_store(
        fx.Vector.from_elements([w0, w1], fx.Int32),
        rsrc(c["xo"]),
        t * (d.q_rows // 4) + (h * HEAD_DIM + d0) // 4,
        cache_modifier=CM_DEV,
    )
    if lane % 4 == 0:
        bo.buffer_store(
            code,
            rsrc(c["xos"]),
            t * (d.q_rows // 32) + (h * HEAD_DIM + d0) // 32,
            cache_modifier=CM_DEV,
        )
    publish(c, c["crdy"], t * hg_n + hg)

stage_split(c, unit)

Flash-decoding over chunk unit % nchunk of token unit // nchunk (split_chunk). A build whose key table is wider than a token's top-k (bounded: every compressed row of a long context) skips a chunk past the token's keys at most, known from its position alone: no load of its q or keys.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def stage_split(c, unit):
    """Flash-decoding over chunk ``unit % nchunk`` of token ``unit // nchunk``
    (``split_chunk``). A build whose key table is wider than a token's top-k
    (``bounded``: every compressed row of a long context) skips a chunk past
    the token's keys at most, known from its position alone: no load of its q
    or keys."""
    nchunk = c["nchunk"]
    t = unit // nchunk
    ch = unit % nchunk
    if const_expr(c.get("bounded", False)):
        d = c["d"]
        pos = ld_i32(c["args"]["pos"], 2 * t)
        cap = fx.min((pos + 1) // c["ratio"], fx.Int32(d.topk_width)) + WINDOW
        if ch * chunk_keys(d, c["S"]) < cap:
            split_chunk(c, t, ch)
    else:
        split_chunk(c, t, ch)

stage_woa(c, task, ops, t0, n)

Rows 32 task .. of wo_a -> bf16 -> vLLM's MXFP8 of the 32 -> X8B.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def stage_woa(c, task, ops, t0, n):
    """Rows 32 task .. of wo_a -> bf16 -> vLLM's MXFP8 of the 32 -> X8B."""
    tid, red, d, lane = c["tid"], c["red"], c["d"], c["lane"]
    split = WOA_SPLIT_K
    gemv_mfmas(c, d.group_k, c["xl"], c["xsl"], red, ops, split, tiled=True, rows=n)
    gpu.barrier()
    mine = tid < WOA_ROWS * n
    tl = fx.min(tid // WOA_ROWS, n - 1)
    t = plus(t0, tl)
    row = tid % WOA_ROWS
    tile = row // ROWS
    y = bf16_round(
        row_sum(red, row % ROWS, tl, [tile * split + w for w in range(split)])
    )
    amax = butterfly(abs(y), (1, 2, 4, 8, 16), fx.max)
    code = mx_code(amax)
    q = clamp_fp8(y * mx_mul(code))
    w = fp8_pack4(
        q,
        *[
            lane_gather(q.bitcast(fx.Int32), fx.min(lane + k, 63)).bitcast(fx.Float32)
            for k in (1, 2, 3)
        ],
    )
    if mine & (row % 4 == 0):
        bo.buffer_store(
            w,
            rsrc(c["x8b"]),
            t * (d.o_rows // 4) + (task * WOA_ROWS + row) // 4,
            cache_modifier=CM_DEV,
        )
    if mine & (row == 0):
        bo.buffer_store(
            code, rsrc(c["x8bs"]), t * (d.o_rows // 32) + task, cache_modifier=CM_DEV
        )
    gpu.barrier()

stage_wob(c, task, ops, task2, ops2, has2, t0, n, mark=False)

Rows 16 task .. (and 16 task2 .., has2) of wo_b -> bf16 -> out, or (c["wob_out"], the mono layer) the caller's hook with each bf16 pair.

Source code in vllm/models/deepseek_v4/amd/mono/attention/back.py

@traced
def stage_wob(c, task, ops, task2, ops2, has2, t0, n, mark=False):
    """Rows 16 task .. (and 16 task2 .., ``has2``) of wo_b -> bf16 -> out, or
    (``c["wob_out"]``, the mono layer) the caller's hook with each bf16 pair."""
    tid, red = c["tid"], c["red"]
    o_rows = c["d"].o_rows
    a = c["args"]
    gemv_mfmas(c, o_rows, c["xl"], c["xsl"], red, ops, rows=n)
    gemv_mfmas(c, o_rows, c["xl"], c["xsl"], red + RED, ops2, rows=n)
    gpu.barrier()
    if const_expr(mark):
        stamp(c, 15)
    per = ROWS * n // 2  # bf16 pairs a task
    k = tid // per
    emit = c.get("wob_out") or (lambda t, col, v0, v1: _wob_store(a, t, col, v0, v1))
    if (tid < 2 * per) & ((k == 0) | has2):
        u = tid % per
        rp = u % (ROWS // 2)
        tl = u // (ROWS // 2)
        t = plus(t0, tl)
        rk = red + k * RED
        col = (k == 0).select(task, task2) * ROWS + 2 * rp
        v0 = row_sum(rk, 2 * rp, tl)
        v1 = row_sum(rk, 2 * rp + 1, tl)
        emit(t, col, v0, v1)
    gpu.barrier()