If you cope with the motors not sharing torque ALL the time, then this (not sharing) may be ok as long as you calibrate the control systems not to fight each other!
Persumably at high loads on the machine, you need both motors to supply enough torque/power to keep the load moving, but you could cal the aggressiveness of the PID so really the second motor only joins in when the load exceeds (or gets close too) the limit of the first motor. If you use a primarily Proportional control, there has to be a speed error at all times (because the Poutut term = error x Pgain), So the level of speed "sag" is going to massively increase after the first motor controller hits it's torque limit (which you could artificially limit to below it's max rating if you were worried about motor thermals).
So if the absolute speed of the system is not that important, you could effectively calibrate the PIDs to not share, and make one motor do most of the general speed keeping and the other motor add torque when required. Again this comes back to one in local speed control, one in local Torque control mode in effect.
Given that you get a feedback of motor torque, you could calibrate the second PID as a much more relaxed (perhaps no I term at all) controller, and then you external controller could actually send it an increasing speed setpoint which would have the effect of increasing it's torque output as it's speed error grows. That speed target could then be calculated by your external master controller based on a long term average motor torque, ie
Say your set point is 1,000rpm (for example)
Send 1,000 rpm target to motor 1, which has I term set to drive machine to 1,000 rpm
Send 1,000rpm target to motor 2, see what amount of torque is is adding as a ratio of the torque being generated by motor 1, and send a higher or lower value to add a relatively slow "balancing" offset. In all cases, both motors are in speed control and so the high speed velocity control is local and in the primary control loop. You will need to low pass filter and average the two feedback torques from motor 1 and motor 2, and have some form of "Gain table" to calculate a suitable adder to the speed target for motor 2
Of course, if you find it enough to have a single drive in speed control, then that long term average "balancing" value can be calulated and applied in the torque domain, ie you take the feedback torque from motor 1, and send it to motor 2, but have a trim value that lets you balance things out as necessary at a slower rate.