GitLab

* Add support for Fused Attn MLA head_dim_qk != head_dim_v
	Modify is_fused_attn_kernel_available() to accept different head_dims for qk and v
	Modify FusedAttnHelper to accept different head_dims for qk and v and modify assert dims checks in parse_qkv_aval()
	Modify FusedAttnFwdPrimitive and FusedAttnBwdPrimitive to accept different head_dims for qk and v
	Modify Fused Attn related cpp and csrc extension API calls to accept different head_dims for qk and v
	Modify DotProductAttention call() to extract head dims separately for qk and v
	Modify the FusedAttn Tests to accommodate for API changes in FusedAttn API
	Add test case for head_dim_qk != head_dim_v (failing)
	Modify the baseline JAX appropriately to reshape the output vector based on v dims and not q dims
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Fix context dims in general DPA in test_fused_attn
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Fix dim for output tensor by replacing with v head dim rather than q head dim
Add test cases for jax fused attn where head_dim_qk != head_dim_v for a combination of data types and attention type
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Modify the fused attn jax unit test case for head dim qk != head dim v
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Use new FusedAttnRunner function signature for separate hidden dim for qk and v in Fused Attn distributed tests
Code clean up
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Fix usage of is_fused_attn signature in distributed tests
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Remove unnecessary assert
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

---------
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* add support for head dim > 128
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* remove debugging
Signed-off-by: Charlene Yang <charleney@nvidia.com>

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

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



* raise tols slightly to tolerate 1/2048 mismatches
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix is_training for test_te_layer
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add bprop support for blackwell
Signed-off-by: Charlene Yang <charleney@nvidia.com>

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

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



* minor tweak for format
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix backend selection results
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* bump sm100 to sm100+
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add sq=1 test for MLA
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* enable sq=1 for bprop
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* minor tweak in comments
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix head_dim logic and remove pytest skip
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add FE fix for d>128
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* update FE again to take in small fixes
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* add cuDNN version info in L0 tests
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* increase tols for Unfused + large dim
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* Revert "add cuDNN version info in L0 tests"

This reverts commit 3e1b426ca5319a2c0540b9e73bba7047d0e583e5.
Signed-off-by: Charlene Yang <charleney@nvidia.com>

* fix tols for Unfused
Signed-off-by: Charlene Yang <charleney@nvidia.com>

---------
Signed-off-by: Charlene Yang <charleney@nvidia.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>

* fixes for jittable grouped_quantize

* fixes for jittable grouped_gemm

* fix contracting_dim for wgrad gemm

* exclude jitted grouped_gemm from the unit test as it does not work cudaGraph

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Implemented GroupedDense and TestGroupedDense for BF16, FP16, and FP8 
* Fix GroupedGemmFFI cuBLAS workspace alignment bug
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Hua Huang <huah@nvidia.com>

Revert "[JAX] GroupedDense v.2 without dynamic shape (#1721)"

This reverts commit 5d01ef21

.
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Implemented GroupedDense and TestGroupedDense for BF16, FP16, and FP8 
* Fix GroupedGemmFFI cuBLAS workspace alignment bug
Signed-off-by: Hua Huang <huah@nvidia.com>
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* refactor the multi_stream utils + implement nvte_multi_tensor_quantize in TE/Common

* implement GroupedQuantizer and grouped_quantize in jaxx

* fix logical_axes_names for transpose tensor in ScaledTensor
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Hua Huang <huah@nvidia.com>
Co-authored-by: Ming Huang <mingh@nvidia.com>

* fix otype for fp8 gemm
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Fix 1x quantize kernel availability check on hopper

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Use 1x quantization + jax transpose on BW for performance
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Use 1x quantization on Hopper as well as it is also faster
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Undo architecture check helper function
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Lint
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Support SWA in CP Ring Attn THD striped sharding
Signed-off-by: Hua Huang <huah@nvidia.com>

* Add some comments; move check to _FusedAttnCPWithP2PHelper.check_supported()
Signed-off-by: Hua Huang <huah@nvidia.com>

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

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



Remove unused check
Signed-off-by: Hua Huang <huah@nvidia.com>

---------
Signed-off-by: Hua Huang <huah@nvidia.com>

* removes unnecessary reshapes for FP8 GEMM

* use nn.jax.scaled_matmul
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* fp8 gemm with direct quant
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

Update jax_scaled_masked_softmax to match TE kernel implementation
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

Make primitive names more granular for better disabling granularity
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

This reverts commit 5bee81e2

.
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* removes unneccessary reshapes for FP8 GEMM

* use nn.jax.scaled_matmul
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Enforce input sharding of norm primitive does not shard hidden dim
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix partitioning issue in dact primitive causing NaN and add better shape checks before calling TE API
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Move dact shape assertion from cpp to python
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* exclude GroupedGemm APIs
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

Fix distributed layernorm test failure
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Update test_helper.py and add QuantizeConfig class for CurrentScaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* WIP distributed current scaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Distributed Current Scaling (debugging). Distributed implementation with replicated scale_inv works for layernorm_mlp but feels like a hack

Has different per-device scale_inv values, but jax.debug.print only shows one of them. Since we're telling JAX/XLA that this scale is replicated, I think it assumes all the values are equal. However, it doesn't actually check this, so it seems we are able to get away with per-device scales for current scaling but I am not sure how stable this will be and may randomly fail if us or the user changes partitioning at all or if XLA decides to actually act on the assumption that all these scale_invs are the same.
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Implement distributed current scaling by computing a global amax and scale before quantization
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add encoder and mnist tests for current scaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add primitive prefix to shardy unique_vars to prevent factor conflicts when performing unfused primitives for current scaling
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove scale_shape primitive arg that is no longer used
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Format
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix expected result on multiprocessing encoder test
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Lint fix
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Update multiprocessing current scaling tolerances
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Uncomment test case that was disabled for testing
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove commented out debug line
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* [JAX-Q] Single GPU current scaling for JAX
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix scale check dtype for MXFP8 scales affecting tests using assert_bitwise_scaled_tensors
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Address comments
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove cast to fp32 for norm primitives now that zero-centered gamma dtype issue is fixed
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix lint issue
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Remove unnecessary cast to fp32
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Lint
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

Check CuDNN version and apply unfused norm if
below a version with the fix
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add a flag to support computing zero-centered gamma in weight dtype or compute dtype for CuDNN
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Address comments
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

---------
Signed-off-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Add test cases for full coverage in jax/test_layer.py
- causal and window size None
- causal and window size default (-1,1)
- no_mask and window size default (-1,1)
- no_mask and window size default (2,2)
- padding and window size None
- padding_causal and window_size (2,2)
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Correct the condition where padding_causal_mask was being mapped to scaled upper triangle
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Fix Issue #1524
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Add a runner and test cases for jax.flax.module.Softmax class for fwd pass only
Segregate runner classes for Softmax module and softmax primitives
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Simplify logic when picking softmax primitives and softmax jax framework calls
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Simplify the logic for performing jax based softmax
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

* Code clean up
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Add support table for mask, SWA and Softmax type. Code linting
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Explicit SWA conditons in comments. Fix Typo
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



* Resolve typo to remove None in SWA comments section
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>

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

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



---------
Signed-off-by: Kshitij Janardan Lakhani <klakhani@nvidia.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

* Add experimental Shardy support.

Production use is not yet recommended.

---------
Signed-off-by: Johannes Reifferscheid <jreiffers@nvidia.com>

* New GroupedGemmPrimitive using variadic args

* Remove squeeze() to reduce D2D memcpy

* Revert to the list append fashion to simplify code

---------
Signed-off-by: Hua Huang <huah@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* scaling enum abstract

* rm NVTE_ from ScalingMode names

* rework scaling mode enum in grouped gemm

* fix norm sharding

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* rm no scaling enum
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* update jax enum
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* rename QuantizeAxis to QuantizeLayout, get_layout to get_data_layout, q_axis to q_layout

* add fatten_axis option

* added gated act to test encoder

* sharding constraint fixes

* fix padding when flattening first dim needs to be padded

* update test sizes so that padding is tested

* rm output sharding as it can be done in the flax module

* sharding scale_inv for mxfp8

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* refactor + mxfp8

* added grouped gemm

* rename linear to dense

* added cublas init phase for groupedGemm

* relax the tol of test encoder multiprocessing mxfp8 by 0.001
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>
Co-authored-by: Hua Huang <huah@nvidia.com>
Co-authored-by: Jeremy Berchtold <jberchtold@nvidia.com>

* Fix softmax shape for THD format.
Signed-off-by: Michael Goldfarb <mgoldfarb@nvidia.com>

Make ffi compatible with jax 0.4
Signed-off-by: Reese Wang <rewang@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

Remove xla_ignore_channel_id check and ignore Scan loop warning in unit test
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix wheel install after src install
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix JAX imports
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* switch order of dirs for finding so
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Use existing dir src build
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Fix lint
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

---------
Signed-off-by: Kirthi Shankar Sivamani <ksivamani@nvidia.com>

* Support THD + ring attention for self attn
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Consolidate reorder strategy
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix dataclass frozen issue
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove redundant code
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use AttnBiasType, AttnMaskType, QKVLayout in cpp_extension
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix lint
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Refine P2P helper check_supported
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add segment_ids/pos check
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fixup
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add dual chunk swap example
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Align different reorder code structure
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* Expose THD to flex MHA module
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Enhance docs
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>
Co-authored-by: Phuong Nguyen <phuonguyen@nvidia.com>

* fixes L1 test

* fix test_multigpu_encoder

* fixes for other multi-encoder tests

* jax.extend.ffi to jax.ffi

* initialization with float32

* add init_dtype as an optional arg to all modules

* update use_scan query from xla flags

* relax threshold for test_encoder fp8

* relax the tols

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* flax module to init params with given dtype
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* all tests passed
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* remove unneccessary reshape for kernel
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* remove casting output of dot
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* clean up
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

---------
Signed-off-by: Phuong Nguyen <phuonguyen@nvidia.com>

* POC for segment_ids/segment_pos
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Change segment_pos position
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use RemainingArgs to solve number of parameters mismatches
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Test mask_descriptor for accomendating different mask representations
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix bugs
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Use descriptor in bwd
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Primitives only accepts pure jnp array
Signed-off-by: Reese Wang <rewang@nvidia.com>

* segment_ids/pos support POC
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Move seqlens/offsets generation to mask descriptor
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Rename MaskDescriptor to SequenceDescriptor
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Generalize get_seqlens_and_offsets
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Utilize sequence desc on FA bwd
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Migrate to new API
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Add docstrings
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove small inputs and test different input format
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix lint
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix seed shardings
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Optimize sequence converting overhead
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Optimize seq_offsets calculation
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix up
Signed-off-by: Reese Wang <rewang@nvidia.com>

* fix lint
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Fix conflicts
Signed-off-by: Reese Wang <rewang@nvidia.com>

* Remove reduntant line
Signed-off-by: Reese Wang <rewang@nvidia.com>

---------
Signed-off-by: Reese Wang <rewang@nvidia.com>