# DeepSeek-V4 NVFP4 Kernel Suite

CuTeDSL kernels for DeepSeek-V4 (Blackwell B200, SM100). All kernels use `cutlass.cute` (CuTeDSL) with Blackwell tensor cores.

## Status (May 21, 2026 — 04:35 UTC)

### ✅ Stage A: Bare Q@K^T via tcgen05.mma → TMEM → GMEM — COMPLETE

**File**: `tests/test_stage_a_v2.py`
**Result**: Q(128,128) @ K^T(128,128) → S(128,128), cosine 0.999999

### 🔨 Stage B: Two MMAs + Identity Softmax — IN PROGRESS

**Pipeline deadlock: FIXED. Kernel runs without deadlock.**
**Bug 1 (V MN-major): Fix applied.**
**Bug 2 (softmax packing): Confirmed correct (V=I test: cosine 1.0).**
**Bug 3 (ACCUMULATE): Fix applied.**
**Bug 4 (non-square PV): PV works for (128,128) output, broken for (128,64) output.**

#### Bug 1: V B-Operand Must Be MN-Major — ✅ FIX APPLIED

V must be shaped (head_dim, seq) = (64, 128) with strides (1, 64) — MN-major.
PV MMA uses `v_major` (OperandMajorMode.MN) instead of `b_major` (K).

V must use `as_strided` — default PyTorch (64,128) gives strides (128,1) which is K-major.

#### Bug 2 (Packing): C-Fragment Composition Store — ✅ CONFIRMED CORRECT

FP32→BF16 packing via C-fragment composition store (FMHA pattern) is correct.
Proven by V=I test (cosine 1.0) and random V 128x128 test (cosine 0.999999).

⛔ **FOOTGUN**: `St32x32bOp` MUST use Float32, NOT BFloat16.
⚠️ The recast view for P packing uses the LOAD layout (128 BF16 elements), not the store composition shape.

#### Bug 3 (ACCUMULATE): First PV Must Use ACCUMULATE=False — ✅ FIX APPLIED

If ACCUMULATE=True on the first PV, `O = P@V + old_O` adds uninitialized TMEM. Always ACCUMULATE=False for first PV, then True for subsequent tiles.

#### Bug 4 (CURRENT): PV MMA Broken for Non-Square Output — 🔨 ROOT CAUSE UNKNOWN

**What works:**
- PV with (128,128) output, V=I: cosine 1.0 ✅
- PV with (128,128) output, random V: cosine 0.999999 ✅

**What doesn't work:**
- PV with (128,64) output, V MN-major (64,128): cosine ~0.01 ❌

**Possible causes:**
1. `make_trivial_tiled_mma` with (128,64) produces different A-fragment layout — alias with softmax P may break
2. V TMA load wrong for (128,64) PV — SMEM layout, TMA descriptor, or partitioning incorrect
3. Epilogue/gC mismatch — output c is (128,64) but epilogue may write (128,128) tile
4. PV mma_tiler_mn doesn't affect the MMA atom (which is always (128,128,16))

**Diagnostic findings:**
- Pointer arithmetic correct: softmax P and PV A-fragment address same TMEM location
- Layout aliasing correct: C-fragment composition and A-fragment produce same physical addresses
- Pipeline ordering correct: softmax completes before PV starts
- Softmax packing correct: proven by V=I test

### 🔨 Stage C: Online Softmax — AFTER B

Per the pseudocode: epilogue warps compute per-row tile_max, rescale, exp, store P back to TMEM.

### 🔨 Stage D: FP8 Paged KV Gather — AFTER C

Replace BF16 TMA load with FP8 paged KV gather + per-position dequant.

---

## Pipeline Deadlock — ✅ FIXED (May 21)

v20-v25 all deadlocked on GPU. Three root causes found and fixed:

### Fix 1: PipelineUmmaAsync for mma_si Must NOT Pass cta_layout_vmnk

FMHA's mma_s0/mma_s1 PipelineUmmaAsync calls do NOT pass cta_layout_vmnk. Removing it fixes the deadlock.

### Fix 2: TMA Warp Must NOT Call tmem.wait_for_alloc()

The tmem allocation barrier has `num_threads = 32 * (mma_warp + epilogue_warps)`. The TMA warp is NOT part of this barrier. Calling `wait_for_alloc()` from the TMA warp corrupts the barrier.

### Fix 3: PipelineTmaStore (not TmaStorePipeline)

`pipeline.TmaStorePipeline` does not exist. The correct name is `pipeline.PipelineTmaStore`.

---

## ⛔ DEAD TEST: test_stage_b_v21.py — DELETED, DO NOT RECREATE

v21 attempted both Bug 1 and Bug 2 fixes in a hand-rolled pipeline kernel. It deadlocks on GPU. Root cause: pipeline synchronization mismatch. **Do not recreate.** Write from scratch using fmha.py as the reference.

---

## ⛔ FOOTGUNS — CUTLASS CuTeDSL Landmines

### 1. St32x32bOp with 16-bit dtype → ILLEGAL MEMORY ACCESS

`St32x32bOp(Repetition(N), BFloat16)` crashes at runtime. You MUST use `St32x32bOp(Repetition(N), Float32)` and pack 2×16-bit values into 1×Float32 backing words via `cute.recast_ptr`. The 16-bit type only appears in the recast view, never in the store atom itself.

### 2. V B-Operand Major Mode ≠ K Major Mode

FMHA requires `v_major_mode == OperandMajorMode.MN`. Passing K's K-major mode for V is WRONG. V must be shaped (head_dim, seq) with strides (1, head_dim) to produce MN-major. Standard PyTorch row-major (seq, head_dim) gives K-major.

### 3. CuTe Nested Layout Modes Flatten Sequentially

A layout like `((128,16),1,(4,2)):((65536,1),0,(16,64))` looks "non-sequential" but flattens to `addr = m*65536 + k` when k = k0 + 16*k1 + 64*k2 (CuTe row-major order). Do NOT assume nested modes imply non-sequential physical addressing. The C-fragment composition and A-fragment alias the same TMEM columns.

### 4. PipelineUmmaAsync Consumer Group = Thread Count, NOT Warp Count

```python
# WRONG: consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 4)
# CORRECT: consumer_group=pipeline.CooperativeGroup(pipeline.Agent.Thread, 32 * len(warp_ids))
```

### 5. PipelineUmmaAsync for mma_si Must NOT Pass cta_layout_vmnk

Passing `cta_layout_vmnk` to the mma_si PipelineUmmaAsync causes deadlock. FMHA does not pass it. Remove it.

### 6. TMA Warp Must NOT Call tmem.wait_for_alloc()

The tmem allocation barrier only includes MMA + epilogue warps. The TMA warp is excluded. Calling `wait_for_alloc()` from the TMA warp corrupts the barrier.

### 7. PV MMA ACCUMULATE Must Be False on First Tile

If ACCUMULATE=True on the first PV MMA, `O = P@V + old_O` adds uninitialized TMEM to the result. Always set ACCUMULATE=False for the first PV, then True for subsequent tiles. FMHA: `pv_tiled_mma.set(tcgen05.Field.ACCUMULATE, kphase_idx != 0)`.

### 8. TMEM Pointer Arithmetic: Offset Units Depend on Pointer Type

When computing PV A-fragment offset from the softmax P offset:
```python
# Softmax store: FP32 pointer + tmem_p0_offset (in FP32 elements)
tStS_P = cute.make_tensor(tStS.iterator + tmem_p0_offset, tStS_P_layout)

# PV A-fragment: BF16 pointer + scaled offset (in BF16 elements)
p_offset = acc_dtype.width // q_dtype.width * tmem_p0_offset  # 2 * 32 = 64
tOrP0 = cute.make_tensor(tOrP.iterator + p_offset, tOrP.layout)
```
Both must address the same physical TMEM column. The 2× scaling accounts for FP32→BF16 element size difference.

---

## Architecture: Per-Tile Flow

```
For each KV tile:
  1. Load warp writes sKV[stage] (paged FP8 gather via indexed cp.async)
  2. MMA warp issues MMA1: sQ @ sKV[stage]^T → tmem_scores (accumulate=False)
     Signals scores_full_mbar (via PipelineUmmaAsync commit)
  3. Epilogue warps wait on mma_si consumer (scores ready), then:
     a. tcgen05.ld scores from TMEM → register fragments
     b. Compute tile_max, new_max, rescale = exp(old_max - new_max)
     c. Apply rescale to tmem_output IN PLACE (tmem_output *= rescale)
     d. tcgen05.st exp(scores - new_max) back to TMEM → P operand (via C-fragment composition)
     e. Release mma_si (softmax_done — MMA warp can re-acquire and issue PV MMA)
  4. MMA warp waits on mma_si acquire (softmax done), MMA2: P @ sV → tmem_output (accumulate=True)
  5. Stage released, load warp can refill it

After all tiles: epilogue warps tcgen05.ld tmem_output, divide by row_sum, cast to BF16, store to GMEM
```

---

## Test Results

| File | Description | Cosine | Status |
|------|-------------|--------|--------|
| `test_stage_a_v2.py` | Q@K^T only | 0.999999 | ✅ PASS |
| `test_mma_si_only.py` | Q@K^T + mma_si pipeline (no PV) | 0.999999 | ✅ PASS |
| `test_softmax_only.py` | Q@K^T + softmax packing, output S | 0.52 | ❌ S overwritten by P (expected) |
| `test_mma_si_pv.py` | Q@K^T + softmax + P@V (V MN-major, 128x64) | 0.01 | ❌ PV output garbage |
| `test_pv_diag.py` | Q@K^T + softmax + P@V (V=I/random, 128x128) | 1.0 / 0.999999 | ✅ PASS |
| `test_layout_compare.py` | Print TMEM layouts for QK S and PV A-fragment | N/A | ℹ️ layout inspection |
| `test_stage_b_v7.py` | Q@K^T + C-fragment softmax (V=K, wrong major) | -0.02 | ❌ wrong major + P packing |
| `test_stage_b_v20.py` | Q@K^T + softmax (V=K, PipelineTmaStore bug) | N/A | ❌ compile error |

---

## Critical APIs & Lessons

### TMEM offset arithmetic
- `find_tmem_tensor_col_offset(fragment)` — returns physical TMEM column count
- QK accumulator: 128 TMEM columns
- A-fragment offset: `acc_dtype.width // q_dtype.width * tmem_p0_offset` (F32/BF16=2)

### pv_mma_tiler — FMHA Convention
```python
pv_mma_tiler = (qk_mma_tiler[0], qk_mma_tiler[2], qk_mma_tiler[1])
# = (M, head_dim, QK_N) = (128, 64, 128) for head_dim=64
```

FMHA passes `pv_mma_tiler[:2] = (128, head_dim)` to `make_trivial_tiled_mma`, NOT the QK tiler `(128, 128)`.

### make_trivial_tiled_mma — Use New Overload
```python
make_trivial_tiled_mma(a_dtype, b_dtype, a_leading_mode, b_leading_mode,
                        acc_dtype, cta_group, mma_tiler_mn, a_source=SMEM)
```

### 3D tensors required
Tensors must be 3D (M, K, L) for `cute.local_tile` — add L=1 dimension.

### Other APIs
1. `cutlass_torch.from_dlpack(t).mark_layout_dynamic(leading_dim=...)` — CuTe tensor from PyTorch
2. `PipelineTmaUmma.create(...).make_participants()` — returns (producer, consumer) pair
3. `utils.gemm.sm100.epilogue_tma_store` — handles transform + partition/dcopy. DO NOT hand-roll.
4. `smem.allocate_tensor()` — for SMEM tensors
5. `LayoutEnum.from_tensor(a).mma_major_mode()` — major mode from cute tensor

## Environment

- **Server**: root@45.76.247.107 (B200, 180 GiB HBM3e per GPU)
- **venv**: `source /root/dsv4-nvfp4-workspace/venv/bin/activate`
- **PYTHONPATH**: `/root/dsv4-nvfp4-workspace/kernel`
- **Model**: `/root/nvidia-meeting/DeepSeek-V4-Pro-NVFP4`
- **vLLM repo**: `/root/dsv4-nvfp4-workspace/vllm` (modified for Blackwell)
- **Pseudocode**: `/root/fragile-kernel-example/README.md`
- **fmha.py reference**: `/root/cutlass/examples/python/CuTeDSL/cute/blackwell/kernel/attention/fmha/fmha.py`
- **fmha_bwd.py reference**: `/root/cutlass/examples/python/CuTeDSL/cute/blackwell/kernel/attention/fmha/fmha_bwd.py`