docs: major README update — packed FP4 SMEM layout, L1 epilogue, TMA descriptors

Added detailed documentation of the packed FP4 architecture:
- mxf4nvf4 reads packed (2 per byte), NOT unpacked like mxf8f6f4
- SMEM layout: float_e2m1_t, BLOCK_K/2 swizzle, UMMA desc byte math
- L1 epilogue: st.shared.u16, no swizzle, kWarpBytesPerRow
- Host TMA: hidden/2 K-dim, block_k/2 inner, fp4_unpacked_smem=false
- Build history through Build 35
This commit is contained in:
2026-05-11 22:40:09 +00:00
parent 30d72e7ef5
commit e608a20dec

View File

@@ -16,7 +16,7 @@ The key requirement: target **`sm_100a`** (not `sm_100`). The `a` suffix enables
block-scaled instructions including `mxf4nvf4`. Targeting plain `sm_100` will produce
"Feature '.scale_vec::4X' not supported on .target 'sm_100f'" errors.
## Kernel Architecture (TARGET)
## Kernel Architecture
```
sm100_fp8_nvfp4_mega_moe_impl
@@ -28,10 +28,77 @@ sm100_fp8_nvfp4_mega_moe_impl
├── kNumSFATmemCols = SF_BLOCK_M / 32 * 4
├── kNumSFBTmemCols = SF_BLOCK_N / 32 * 4
├── kNumSFUint32 = kHidden / 64 (4 UE4M3 per int32)
├── UE4M3 L1 epilogue (float → cutlass::float_e4m3_t cast, sign bit cleared)
├── UE4M3 L1 epilogue (float → e4m3 cast, sign bit cleared)
└── recipe = (1, 1, 16)
```
## Critical: mxf4nvf4 Requires FP4×FP4 (Packed)
**The `mxf4nvf4` PTX instruction requires BOTH A and B to be FP4 (E2M1 packed, 2 per byte).**
Unlike `mxf8f6f4` which reads FP4 from SMEM as if it were FP8 (1 byte/element, low nibble=value,
high nibble=ignored via `float_e2m1_unpacksmem_t`), `mxf4nvf4` reads FP4 **packed** from SMEM:
- 2 E2M1 values per byte (low nibble=first, high nibble=second)
- Advances K by 32 bytes per UMMA_K=64 atom (64 packed nibbles = 32 bytes)
- Byte stride of `BLOCK_K / 2` per K-row (not `BLOCK_K * sizeof(dtype_t)`)
### SMEM Layout for mxf4nvf4
```cpp
using a_dtype_t = cutlass::float_e2m1_t; // packed, 4 bits/element
using b_dtype_t = cutlass::float_e2m1_t;
static_assert(cutlass::sizeof_bits_v<a_dtype_t> == 4,
"mxf4nvf4 requires packed FP4 (4 bits/element) in SMEM");
constexpr uint32_t kSwizzleAMode = BLOCK_K / 2; // 64 bytes (packed)
constexpr uint32_t kSwizzleBMode = BLOCK_K / 2; // 64 bytes (packed)
constexpr uint32_t SMEM_A_SIZE_PER_STAGE = LOAD_BLOCK_M * BLOCK_K / 2; // packed
constexpr uint32_t SMEM_B_SIZE_PER_STAGE = LOAD_BLOCK_N * BLOCK_K / 2; // packed
```
### UMMA Descriptors
The `make_umma_desc` and `advance_umma_desc_lo` helpers use `sizeof(dtype_t)` for byte math.
Since `sizeof(float_e2m1_t) == 1` but the real stride is 4 bits/element, pass the effective
byte-stride parameters:
```cpp
// Use BLOCK_K/2 as the template param and uint8_t as dtype for correct byte math
auto a_desc = make_umma_desc<K, LOAD_BLOCK_M, BLOCK_K/2, kSwizzleAMode, false, uint8_t>(...);
// Advance by UMMA_K/2 bytes per atom (64 packed elements = 32 bytes)
a_desc.lo = advance_umma_desc_lo<K, LOAD_BLOCK_M, kSwizzleAMode, uint8_t>(..., k * (UMMA_K / 2));
```
### L1 Epilogue (Packed FP4 Output)
The L1 epilogue writes packed E2M1 to SMEM using direct `st.shared.u16` (no STSM, no swizzle for v1):
```
Each lane quantizes 4 BF16 → 4 E2M1 nibbles → 2 bytes (uint16)
Lane mapping: row_in_atom = lane_idx / 4, col_pair = lane_idx % 4
Row stride: L1_OUT_BLOCK_N / 2 bytes (packed)
kWarpBytesPerRow = L1_OUT_BLOCK_N / 8
```
TMA store: `block_n / 4` inner dim, no swizzle (`CU_TENSOR_MAP_SWIZZLE_NONE` for v1).
### Host-side TMA Descriptors
All activation TMA descriptors use packed FP4 dimensions:
- K-dim: `hidden / 2` bytes (not `hidden` elements)
- Inner block: `block_k / 2` bytes
- Swizzle mode: `swizzle_acts_mode / 2`
- `fp4_unpacked_smem = false` for all activation descriptors
- L1 output: `block_n / 4` inner, swizzle=0 (no swizzle)
### Pybind Buffer Sizing
The `NVFP4SymmBuffer` uses packed byte counts:
- `fp4_token_layout = layout::Data(hidden / 2)` (packed bytes per token)
- `fp4_intermediate_token_layout = layout::Data(intermediate_hidden / 2)`
- Tensor shapes: `{M, hidden / 2}` for uint8 packed activations
## Weight Transformation Pipeline
```
@@ -63,27 +130,12 @@ native UE4M3 scales with block16 grouping.
| TMEM SF cols (SFA) | `SF_BLOCK_M / 32` | `SF_BLOCK_M / 32 * 4` |
| UTCCP col stride | `i * 4` | `i * 8` |
| `kNumSFUint32` | `kHidden / 128` | `kHidden / 64` |
| L1 epilogue | UE8M0 (`>> 23`) | UE4M3 (float→e4m3 cast) |
| SMEM dtype | `float_e2m1_unpacksmem_t` (1 byte/elem) | `float_e2m1_t` (packed, 4 bits/elem) |
| SMEM row stride | `BLOCK_K * sizeof(dtype_t)` = 128 | `BLOCK_K / 2` = 64 (packed) |
| UMMA_K | 32 | 64 |
| L1 epilogue | STSM + UE8M0 scales | st.shared.u16 + UE4M3 scales, no swizzle |
| recipe | `(1, 1, 32)` | `(1, 1, 16)` |
## Critical Implementation Details
### scale_format_ constraint
The CUTLASS instruction descriptor has a single `scale_format_` bit (0=E4M3, 1=E8M0)
that applies to BOTH A and B scale factors. For NVFP4 (E4M3), both activation (SFA)
and weight (SFB) scales must use UE4M3. The L1 epilogue outputs UE4M3 activation scales
(float → `cutlass::float_e4m3_t` with sign bit cleared).
### Arch flag
The JIT compiler MUST target `sm_100a`, not `sm_100`. Without the `a` suffix, the
`mxf4nvf4` instruction is unavailable and compilation will fail with
"Feature '.scale_vec::4X' not supported on .target 'sm_100f'".
### Weight scale_2 folding
The NVFP4 checkpoint has dual-level scaling: per-block UE4M3 + per-tensor float32.
The `weight_scale_2` must be folded into the block scales before packing:
`effective_scale = block_scale * global_scale`, then re-quantize to UE4M3.
## Build History
| Build | Error | Fix |
@@ -101,14 +153,19 @@ The `weight_scale_2` must be folded into the block scales before packing:
| 19 | kGranK still 16 in C++ binding | Should stay 16 — was wrongly changed to 32 |
| 20 | `uint32 >> 23` fails | Cast to int32 first |
| 22 | Garbled output | Fell back to mxf8f6f4 — should use mxf4nvf4 on sm_100a |
| 2324 | transform_nvfp4 l1_weight_scale_2 | Added global scale folding to Python |
| 25 | Triton float8→uint8 cast | Manual FP32→E4M3 bit manipulation |
| 25 | Triton `__rpow__` | IEEE 754 bit-level 2^exp |
| 2628 | Unpacked SMEM (wrong for mxf4nvf4) | Half-zeroed accumulators |
| 2931 | Packed FP4 SMEM + STSM | Wrong: STSM writes 1 byte/elem, MMA reads 2/elem |
| **32** | **Full packed FP4 revert** | **float_e2m1_t, BLOCK_K/2 swizzle, UMMA desc fixes** |
| 33 | Syntax error in patch | Orphan `@triton.jit` decorator |
| 34 | Triton nested `tl.where` | Sum-of-comparisons for E2M1 quantization |
| 35 | Triton `constexpr[0]` indexing | `tl.split()` for E2M1 pair packing |
## Remaining Work
- [ ] Fix DeepGEMM JIT to target `sm_100a` instead of `sm_100`
- [ ] Add NVFP4 MMA kind enum to DeepGEMM runtime (not just MXFP8FP4 with NVFP4 hat)
- [ ] Revert to Build 17's `mxf4nvf4.scale_vec::4X` instruction (was correct, just wrong arch)
- [ ] Revert `kGranK` to 16, UE4M3 scales, block16 SF layout
- [ ] Add `get_sf_uttcp_aligned_block_sizes` branch for block16 layout
- [ ] Remove UE4M3→UE8M0 conversion and block16→block32 merge from Python
- [ ] Verify TMEM 4X layout (i*8 stride, 4 sub-columns)
- [ ] End-to-end quality test on B200
- [ ] End-to-end quality test on B200 (Build 35+)
- [ ] Add 32B/64B swizzle to L1 epilogue for perf (v2)
- [ ] Optimize L1→L2 bandwidth with packed format
- [ ] Performance benchmarking vs standard FusedMoE path