uki2 ddlZddlmZddlmZddlmZddlmZddlm Z ddl m Z dZ d Z ejd Zd e_ej#e ej%e d Zd Zej*eeddZy)N)api)core) tree_util)mlir)ir)hloc4~tj||SN)jax_core jaxpr_as_funjaxprargs unused_kwargss P/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/jax/_src/cudnn/fusion.py_cudnn_fusion_implrs %  u %t ,,c~~|jSr ) out_avalsr s r_custom_abstract_evalrs r cudnn_fusionTc$tjtj|fi|d|\}}t j |}t j |}tj||j|d}t j||S)z`Creates a new cudnn_fusion corresponding to calling the given function f with args and kwargs.T) return_shape)namer) r make_jaxpr functoolspartialr tree_leavestree_structurecudnn_fusion_pbind__name__tree_unflatten)frkwargsr out_shapes flat_argsout_treeout_flats rcall_cudnn_fusionr**scnn a"6"  %##D))  % %j 1(  )!**E J(  ! !(H 55rc`tj|g||dz|d}|dj}|jd}t j |j Dcgc]}|jc}|jdtjj|g}|j Scc}w)zMake cudnn_fusion which calls the implementation function. Currently this leaks a CallOp since we're using the `core_call_lowering` function, but this should get cleaned up by DCE easily. z.impl)r call_jaxprrcalleez__cudnn$fusion)call_target_namecalled_computations) rcore_call_loweringowner attributesr CustomCallOpresultstypeoperandsr ArrayAttrget) ctxrrrimplcall_op called_fnrrs r _cudnn_fusion_stablehlo_loweringr>6s     G^ $ GMM'  *)!!__%QVV% % (()5 ,    &sB+cuda)platformc6tjt|S)a>Makes a function become a cuDNN kernel. Relies on XLA's handling of custom fusions with __cudnn$fusion backend. Currently limited to GEMM fusions. For example - batch matmul with mixed types and addition: @cudnn_fusion def fn(x, y, z): return jnp.float32(jax.lax.batch_matmul(jnp.bfloat16(x), y)) + z )rrr*)r$s rrrNs   ,a 00r)rjax._srcrrr rjax._src.interpretersrjax._src.lib.mlirrjax._src.lib.mlir.dialectsrrr Primitiver multiple_resultsdef_abstract_evaldef_implr*r>register_loweringrrrrLs%& *-  $##N3"& !67*+ 6&4v 1r