Eltwise / Reduce / Broadcast Related tasks/bugs - MCW

[umadevimcw]

• ttnn.repeat throws runtime memory error #6361
• Softplus #5426
• ttl.tensor.reduce needs to be run twice to produce correct output #4384
• Add support for broadcast multiply along batch dimension (W) #5769
• Supporting sharded and in-place input & output for tt_lib.tensor.silu & ttnn.silu #6271
• ttnn.l1_loss fails with low PCC on both GS and WH [Bug Report][GS][WH] #6390
• ttl.tensor.rmsnorm results in a mismatch #4327
• Runtime error: ttnn.add fails with tensor on host #4899
• ttnn.lerp operation is missing weight parameter #5817
• logit operation fails with low PCC on Wormhole and Grayskull cards #4405
• [Feature Request] Improvement Needed for Unit Tests #6633
• [Bug Report] invalid asin result #6722
• [Bug Report] invalid acos result #6723
• [Bug Report] invalid isclose result #6672
• [Bug Report] invalid isneginf result #6727
• [Bug Report] invalid isfinite result #6724
• [Bug Report] invalid isposinf result #6726
• [Bug Report] invalid erfc result #6737
• [Bug Report] invalid signbit result #6728
• [Bug Report] Invalid recip result #6720
• ttl.tensor.recip operation breaks with low PCC [Wormhole] [Doc] #3561
• ttl.tensor.recip operation breaks with low PCC [Grayskull] #2898
• [Bug Report] invalid acos result #6723
• [Bug Report] Invalid log result  #6721
• ttnn.exp, exp2, expm1 and ldexp fail with low PCC on both GS and WH in larger value ranges [Bug Report][GS][WH] #6391
• Low PCC failure : backward sqrt - tt_lib.tensor.sqrt_bw [Grayskull] #5148
• tt_lib.tensor.rdiv fails with low PCC [Grayskull] #3310
• ttl.tensor.addcdiv operation fails with low PCC error on WH #3417
• tt_lib.tensor.complex operations fail with low PCC when last input dimension is relatively large #5519
• tt_lib.tensor.complex operations throw RuntimeError for many input dimensions #5518
• Low PCC failure : backward clamp ops - tt_lib.tensor.clamp_max_bw and tt_lib.tensor.clamp_min_bw [Grayskull][Wormhole] #5147
• tt_lib logical operations - Low PCC when one of the operands is BFLOAT_8B [Grayskull and Wormhole] #4673
• tt_lib eltwise_logaddexp fails with low PCC when at least one operand BFLOAT8_B [Grayskull and Wormhole] #4711
• ttnn.logaddexp and logaddexp2 fail with BFLOAT8_B operands [Bug Report][GS][WH]  #6392
• ttnn.geglu operation fails with "Chunk not divisible by tile" error [Grayskull and Wormhole] #6677
• ttnn.reglu operation fails with "Chunk not divisible by tile" error [Grayskull and Wormhole] #6687
• ttnn.glu operation has several issues #6537
• Update TTNN documentation for atan2, hypot, squared_difference #7296
• ttl.tensor.rsqrt operation breaks with low PCC on some ranges [Grayskull] #3521
• ttl.tensor.rsqrt operation breaks with low PCC on some ranges [Wormhole] #3523
• tt_lib.logical_ori Low PCC failure [Grayskull and Wormhole]  #4674
• ttl.tensor.std_hw operation breaks with low PCC [Grayskull] #3015
• ttl.tensor.std_hw operation breaks with low PCC [Wormhole] #3605
• ttl.tensor.var_hw operation breaks with low PCC [Grayskull] #3014
• ttl.tensor.var_hw operation breaks with low PCC [Wormhole] #3604
• ttnn.std and ttnn.var operations failed with low PCC value error [Grayskull and Wormhole] #7052

[jliangTT]

i am subscribe already. thanks

[VirdhatchaniKN]

Regarding Issue #4384 ttl.tensor.reduce needs to be run twice to produce correct output

• We worked on it, and found out that this is specific to dim -2 (Height).
• Tried implementing mean in TT_DNN using reduce, sum, mean_hw , but still facing assertion issue.

Documented my work in detail : Doc Link
Conversation Link in Issue

[umadevimcw]

@jliangTT @tt-aho @eyonland would like get inputs on following items

We have a few bugs that return low PCC due to input range and datatype for example

• ops that depend on exp, will return low PCC after a certain range of inputs
• what could be the maximum input range (low, high value in data gen) that we can use to test for such ops, and also for the data types like bfloat16 and bfloat8?

We have bugs related to storing nan and inf -> how to address those issues?
#4409 #6583 - issues were not able to store nan in WHB0

[umadevimcw]

Regarding Issue #4384 ttl.tensor.reduce needs to be run twice to produce correct output

• We worked on it, and found out that this is specific to dim -2 (Height).
• Tried implementing mean in TT_DNN using reduce, sum, mean_hw , but still facing assertion issue.

Documented my work in detail : Doc Link Conversation Link in Issue

@tt-aho @eyonland @jliangTT
Also need some inputs to proceed further for reduce sum as we are getting low PCC in reduce

[VirdhatchaniKN]

Hi @tt-aho , @eyonland , @jliangTT
Fix provided for three of the above listed issues

• Recip Issue : Fix
• Erfc Issue : Fix
• Addcdiv Issue : Fix

Kindly review, Thanks.