[PyTorch] Add FP8 attention with current scaling (#2012)

* debug existing usage
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix fp8_dpa
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reimplement fp8_dpa
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* more clean up
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update FE develop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* redesign CS; need cleanup
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up s/dP quantizers
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* return dP to DS
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* improve quantizer_helper; tweak dP DS/CS logic
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* debug CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* update FE commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up non-CP; debug dq/dk mismatches
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor success with CP; need to remove debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove debug info
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* disable fp8 output for fp8_mha + CS
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add output_tensor_type to FADescriptor
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes for CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove print
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* more fixes for non-CP and CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* enable non-determinism for blackwell
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix indent; remove print
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* switch from create_tensor_from_data to make_like
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* enable a2a+p2p for CS CP and require additional cp_group_global
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* condense tests; only create dist groups once
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* consolidate CP P2P per-tile calls for fwd/bwd and fused/flash
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix flash-attn from last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes for previous commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix attn_mask_type in f16 causal
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert bb6a0a59 temporarily
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reenable comparison for some tensors in CP tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix dbias for fused attn CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clean up prints/comments and add back NVTE_CS_dP_SCALE
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* first attempt at mixed DS/CS reduction
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fix for last commit for mixed DS/CS reduction
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove prints from 69639024
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix DS recipe for dP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add NVTE_DPA_FORCE_DS to force DS for all DPA tensors, not just dP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix NVTE_DPA_FORCE_DS and add NVTE_PRINT
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* modify DS recipe for MLPerf
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reduce only over TP group; need to think about CP group later
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* streamline fake_recipe/quantizer generation; allow NVTE_DPA_Fixed_Scales or DS-update S/dP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* add more print: NVTE_LAYER_NUMBER
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* split S/dP in env vars: NVTE_DPA_Fix_S_Scale and NVTE_DPA_Fix_dP_Scale
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix autocast_key for DS
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add NVTE_REPEAT_in_F16 to repeat FP8 fwd/bwd passes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add FP8 CS to UnfusedDPA; unsuccessful; does not affect other backends
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* temporary: print min/max and save tensors for debugging
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* emulate q/dq+bf16 with NVTE_Emulate_in_F16; add NVTE_DPA_FORCE_MXFP8 for MXFP8 q/dq
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add RHT to BMM1 with NVTE_RHT_BMM1 for the size
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* re-enable fused attn in dpa_fp8_vs_f16 test; changed during unfused attn implementation
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add NVTE_FP8_CS_POWER_OF_2, NVTE_DPA_FORCE_BLOCKFP8, NVTE_Emulate_QDQ_QKV, NVTE_Emulate_QDQ_O, NVTE_Emulate_QDQ_dO
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add F16 O support for FP8 kernels
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert to TE FE commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* return to FE develop
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* tidy up; untested
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fix for last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes and improvements for last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* more minor fixes and improvements
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* more small fixes/improvements; mostly for CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix CS/DS recipe switch in DPA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* avoid quantizing/saving of O when CS bwd uses F16 O
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* move fp8_autocast(fp8_recipe) print to utils.py
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add debug logging to unit tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add back prints of quantizers/layer_number for debugging
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* enable amax reduction for both CS and DS tensors
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix NVTE_FP8_DPA_BWD=0 for CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix last commit for F16 fwd/bwd a2a+p2p
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* small fixes for float8_current_scaling(), nominal types, and unruly d_out types
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix fp8_output in MHA and some CP tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fixes to CP tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fixes for CP A2A
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* clamp input data in tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove rmse and tighten atol/rtol for tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* restructure fp8_recipes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix linter
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* Revert "remove rmse and tighten atol/rtol for tests"

This reverts commit 15dba6a59a5323d414f02cf22f099cb00d880532.
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* more fixes for linter
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix fp8 recipe changes for F16 code path
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert to FE on main to help with merges
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* switch back to FE develop after merge
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update FE develop commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix last merge
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert to GitHub FE 1.14.1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* update FE to its latest main
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fix for A2A
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix last commit for A2A DS
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove memset for BSHD/SBHD FP8
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* remove concat for qkv quantization in CS
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* improve/simplify the logic for last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add nominal_type for UnfusedDPA FP8 EmuFunc
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* WIP: update env vars for DPA recipes
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix typo in last commit
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix DS recipe creation for NVFP4 global recipe
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace python max with torch.maximum
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix linter
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix CP A2A for FA
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reduce prints in print_quantizers
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add FP8 env vars to NVTE_DEBUG prints
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* add reduce_amax to DS repr
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* separate fp8_dpa/fp8_mha in CP tests; fix A2A for them; add f16_O tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* address some reciews
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* make data optional in create_hp_tensor_with_amax
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* minor fix for comments in bwd
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* print cudnn version in attn tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* disable CS for Hopper
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* alternative tests to reduce CI time
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* make NVTE_DPA_FP8CS_O_in_F16 default to 1
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove _fp8 variables to avoid confusion
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* return to requiring two cp_groups for a2a+p2p
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* replace NVTE_PRINT with NVTE_DEBUG/_LEVEL for quantizer prints
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* provide a basic set of tests for CP
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* fix the last merge with nvfp4 PR
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* disable for Hopper
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix fp8 backend selection for Hopper
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* reduce CP CI to essential tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci



* minor fix to CP test
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* fix recipe logic in tests
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* revert to concat for qkv quantization
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

* remove cudnn version in qa scripts
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>

---------
Signed-off-by: Charlene Yang <8636796+cyanguwa@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

transformer_engine/pytorch/tensor/float8_tensor.py

@@ -215,6 +215,8 @@ class Float8CurrentScalingQuantizer(Quantizer):
     amax: torch.Tensor
     """FP8 datatype"""
     dtype: TE_DType
+    """amax update options"""
+    use_existing_amax: bool
     """amax reduction options"""
     with_amax_reduction: bool
     amax_reduction_group: Optional[dist_group_type]
@@ -229,6 +231,7 @@ class Float8CurrentScalingQuantizer(Quantizer):
         *,
         rowwise: bool = True,
         columnwise: bool = True,
+        use_existing_amax: bool = False,
         with_amax_reduction: bool = False,
         amax_reduction_group: Optional[dist_group_type] = None,
         force_pow_2_scales: bool = False,
@@ -238,6 +241,7 @@ class Float8CurrentScalingQuantizer(Quantizer):
         self.scale = torch.empty(1, dtype=torch.float32, device=device)
         self.amax = torch.empty(1, dtype=torch.float32, device=device)
         self.dtype = fp8_dtype
+        self.use_existing_amax = use_existing_amax
         self.with_amax_reduction = with_amax_reduction
         self.amax_reduction_group = amax_reduction_group
         self.force_pow_2_scales = force_pow_2_scales